K IMI K2.5: V ISUAL AGENTIC I NTELLIGENCE
T ECHNICAL R EPORT OF K IMI K2.5
Kimi Team
A BSTRACT
We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general
arXiv:2602.02276v1 [cs.CL] 2 Feb 2026
agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modali-
ties enhance each other. This includes a series of techniques such as joint text-vision pre-training,
zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foun-
dation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that
dynamically decomposes complex tasks into heterogeneous sub-problems and executes them con-
currently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various
domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency
by up to 4.5× over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint1
to facilitate future research and real-world applications of agentic intelligence.
Figure 1: Kimi K2.5 main results.
1 Introduction
Large Language Models (LLMs) are rapidly evolving toward agentic intelligence. Recent advances, such as GPT-
5.2 [41], Claude Opus 4.5 [6], Gemini 3 Pro [20], and Kimi K2-Thinking [1], demonstrate substantial progress in
agentic capabilities, particularly in tool calling and reasoning. These models increasingly exhibit the ability to decom-
pose complex problems into multi-step plans and to execute long sequences of interleaved reasoning and actions.
1 https://huggingface.co/moonshotai/Kimi-K2.5
Kimi K2.5 T ECHNICAL R EPORT
In this report, we introduce the training methods and evaluation results of Kimi K2.5. Concretely, we improve the
training of K2.5 over previous models in the following two key aspects.
Joint Optimization of Text and Vision. A key insight from the practice of K2.5 is that joint optimization of text and
vision enhances both modalities and avoids the conflict. Specifically, we devise a set of techniques for this purpose.
During pre-training, in contrast to conventional approaches that add visual tokens to a text backbone at a late stage [8,
21], we find early vision fusion with lower ratios tends to yield better results given the fixed total vision-text tokens.
Therefore, K2.5 mixes text and vision tokens with a constant ratio throughout the entire training process.
Architecturally, Kimi K2.5 employs MoonViT-3D, a native-resolution vision encoder incorporating the NaViT packing
strategy [15], enabling variable-resolution image inputs. For video understanding, we introduce a lightweight 3D ViT
compression mechanism: consecutive frames are grouped in fours, processed through the shared MoonViT encoder,
and temporally averaged at the patch level. This design allows Kimi K2.5 to process videos up to 4 × longer within
the same context window while maintaining complete weight sharing between image and video encoders.
During post-training, we introduce zero-vision SFT—text-only SFT alone activates visual reasoning and tool use.
We find that adding human-designed visual trajectories at this stage hurts generalization. In contrast, text-only SFT
performs better—likely because joint pretraining already establishes strong vision-text alignment, enabling capabilities
to generalize naturally across modalities. We then apply joint RL on both text and vision tasks. Crucially, we find
visual RL enhances textual performance rather than degrading it, with improvements on MMLU-Pro and GPQA-
Diamond. This bidirectional enhancement—text bootstraps vision, vision refines text—represents superior cross-
modal alignment in joint training.
Agent Swarm: Parallel Agent Orchestration. Most existing agentic models rely on sequential execution of tool
calls. Even systems capable of hundreds of reasoning steps, such as Kimi K2-Thinking [1], suffer from linear scaling of
inference time, leading to unacceptable latency and limiting task complexity. As agentic workloads grow in scope and
heterogeneity—e.g., building a complex project that involves massive-scale research, design, and development—the
sequential paradigm becomes increasingly inefficient.
To overcome the latency and scalability limits of sequential agent execution, Kimi K2.5 introduces Agent Swarm, a
dynamic framework for parallel agent orchestration. We propose a Parallel-Agent Reinforcement Learning (PARL)
paradigm that departs from traditional agentic RL [2]. In addition to optimizing tool execution via verifiable rewards,
the model is equipped with interfaces for sub-agent creation and task delegation. During training, sub-agents are
frozen and their execution trajectories are excluded from the optimization objective; only the orchestrator is updated via
reinforcement learning. This decoupling circumvents two challenges of end-to-end co-optimization: credit assignment
ambiguity and training instability. Agent Swarm enables complex tasks to be decomposed into heterogeneous sub-
problems executed concurrently by domain-specialized agents, transforming task complexity from linear scaling to
parallel processing. In wide-search scenarios, Agent Swarm reduces inference latency by up to 4.5× while improving
item-level F1 from 72.8% to 79.0% compared to single-agent baselines.
Kimi K2.5 represents a unified architecture for general-purpose agentic intelligence, integrating vision and language,
thinking and instant modes, chats and agents. It achieves strong performance across a broad range of agentic and
frontier benchmarks, including state-of-the-art results in visual-to-code generation (image/video-to-code) and real-
world software engineering in our internal evaluations, while scaling both the diversity of specialized agents and the
degree of parallelism. To accelerate community progress toward General Agentic Intelligence, we open-source our
post-trained checkpoints of Kimi K2.5, enabling researchers and developers to explore, refine, and deploy scalable
agentic intelligence.
2 Joint Optimization of Text and Vision
Kimi K2.5 is a native multimodal model built upon Kimi K2 through large-scale joint pre-training on approximately
15 trillion mixed visual and text tokens. Unlike vision-adapted models that compromise either linguistic or visual
capabilities, our joint pre-training paradigm enhances both modalities simultaneously. This section describes the
multimodal joint optimization methodology that extends Kimi K2 to Kimi K2.5.
2.1 Native Multimodal Pre-Training
A key design question for multimodal pre-training is: Given a fixed vision-text token budget, what is the optimal
vision-text joint-training strategy. Conventional wisdom [8, 21] suggests introducing vision tokens predominantly in
the later stages of LLM training at high ratios (e.g., 50% or higher) should accelerate multimodal capability acquisition,
treating multimodal capability as a post-hoc add-on to linguistic competence.
2
Kimi K2.5 T ECHNICAL R EPORT
Table 1: Performance comparison across different vision-text joint-training strategies. Early fusion with a lower vision
ratio yields better results given a fixed total vision-text token budget.
Vision Injection Vision-Text Vision Vision Text Text
Timing Ratio Knowledge Reasoning OCR Knowledge Reasoning Code
Early 0% 10%:90% 25.8 43.8 65.7 45.5 58.5 24.8
Mid 50% 20%:80% 25.0 40.7 64.1 43.9 58.6 24.0
Late 80% 50%:50% 24.2 39.0 61.5 43.1 57.8 24.0
However, our experiments (as shown in Table 1 Figure 9) reveal a different story. We conducted ablation studies
varying the vision ratio and vision injection timing while keeping the total vision and text token budgets fixed. To
strictly meet the targets for different ratios, we pre-trained the model with text-only tokens for a specifically calculated
number of tokens before introducing vision data. Surprisingly, we found that the vision ratio has minimal impact on
final multimodal performance. In fact, early fusion with a lower vision ratio yields better results given a fixed
total vision-text token budget. This motivates our native multimodal pre-training strategy: rather than aggressive
vision-heavy training concentrated at the end, we adopt a moderate vision ratio integrated early in the training process,
allowing the model to naturally develop balanced multimodal representations while benefiting from extended co-
optimization of both modalities.
2.2 Zero-Vision SFT
Pretrained VLMs do not naturally perform vision-based tool-calling, which poses a cold-start problem for multimodal
RL. Conventional approaches address this issue through manually annotated or prompt-engineered chain-of-thought
(CoT) data [8], but such methods are limited in diversity, often restricting visual reasoning to simple diagrams and
primitive tool manipulations (crop, rotate, flip).
An observation is that high-quality text SFT data are relatively abundant and diverse. We propose a novel approach,
zero-vision SFT, that uses only text SFT data to activate the visual, agentic capabilities during post-training. In this
approach, all image manipulations are proxied through programmatic operations in IPython, effectively serving
as a generalization of traditional vision tool-use. This "zero-vision" activation enables diverse reasoning behaviors,
including pixel-level operations such as object size estimation via binarization and counting, and generalizes to visually
grounded tasks such as object localization, counting, and OCR.
Figure 2 illustrates the RL training curves, where the starting points are obtained from zero-vision SFT. The results
show that zero-vision SFT is sufficient for activating vision capabilities while ensuring generalization across modal-
ities. This phenomenon is likely due to the joint pretraining of text and vision data as described in Section 2.1.
Compared to zero-vision SFT, our preliminary experiments show that text-vision SFT yields much worse performance
on visual, agentic tasks, possibly because of the lack of high-quality vision data.
2.3 Joint Multimodal Reinforcement Learning (RL)
In this section, we describe the methodology implemented in K2.5 that enables effective multimodal RL, from
outcome-based visual RL to emergent cross-modal transfer that enhances textual performance.
Outcome-Based Visual RL Following the zero-vision SFT, the model requires further refinement to reliably incor-
porate visual inputs into reasoning. Text-initiated activation alone exhibits notable failure modes: visual inputs are
sometimes ignored, and images may not be attended to when necessary. We employ outcome-based RL on tasks that
explicitly require visual comprehension for correct solutions. We categorize these tasks into three domains:
• Visual grounding and counting: Accurate localization and enumeration of objects within images;
• Chart and document understanding: Interpretation of structured visual information and text extraction;
• Vision-critical STEM problems: Mathematical and scientific questions filtered to require visual inputs.
Outcome-based RL on these tasks improves both basic visual capabilities and more complex agentic behaviors. Ex-
tracting these trajectories for rejection-sampling fine-tuning (RFT) enables a self-improving data pipeline, allowing
subsequent joint RL stages to leverage richer multimodal reasoning traces.
3
Kimi K2.5 T ECHNICAL R EPORT
MMMU Pro MathVision
0.76 0.78
0.75 0.76
0.74
Accuracy Accuracy
0.74
0.73
0.72
0.72
0.70
0.71
0.68
RL flops RL flops
CharXiv(RQ) OCRBench
0.78 0.92
0.76 0.90
0.74
0.88
0.72
Accuracy Accuracy
0.70 0.86
0.68 0.84
0.66 0.82
0.64 0.80
0.62
0.78
RL flops RL flops
Figure 2: Vision RL training curves on vision benchmarks starting from minimal zero-vision SFT. By scaling vision
RL FLOPs, the performance continues to improve, demonstrating that zero-vision activation paired with long-running
RL is sufficient for acquiring robust visual capabilities.
Table 2: Cross-Modal Transfer: Vision RL Improves Textual Knowledge
Benchmark Before Vision-RL After Vision-RL Improvement
MMLU-Pro 84.7 86.4 +1.7
GPQA-Diamond 84.3 86.4 +2.1
LongBench v2 56.7 58.9 +2.2
Visual RL Improves Text Performance To investigate potential trade-offs between visual and textual performance,
we evaluated text-only benchmarks before and after visual RL. Surprisingly, outcome-based visual RL produced mea-
surable improvements in textual tasks, including MMLU-Pro (84.7% → 86.4%), GPQA-Diamond (84.3% → 86.4%),
and LongBench v2 (56.7% → 58.9%) (Table 2). Analysis suggests that visual RL enhances calibration in areas re-
quiring structured information extraction, reducing uncertainty on queries that resemble visually grounded reasoning
(e.g., counting, OCR). These findings indicate that visual RL can contribute to cross-modal generalization, improving
textual reasoning without observable degradation of language capabilities.
Joint Multimodal RL Motivated by the finding that robust visual capabilities can emerge from zero-vision SFT paired
with vision RL—which further enhances general text abilities—we adopt a joint multimodal RL paradigm during Kimi
K2.5’s post-training. Departing from conventional modality-specific expert divisions, we organize RL domains not
by input modality but by abilities—knowledge, reasoning, coding, agentic, etc. These domain experts jointly learn
from both pure-text and multimodal queries, while the Generative Reward Model (GRM) similarly optimizes across
heterogeneous traces without modality barriers. This pardaigm ensures that capability improvements acquired through
either textual or visual inputs inherently generalize to enhance related abilities across the alternate modality, thereby
maximizing cross-modal capability transfer.
3 Agent Swarm
The primary challenge of existing agent-based systems lies in their reliance on sequential execution of reasoning and
tool-calling steps. While this structure may be effective for simpler, short-horizon tasks, it becomes inadequate as the
complexity of the task increases and the accumulated context grows. As tasks evolve to contain broad information
gathering and intricate, multi-branch reasoning, sequential systems often encounter significant bottlenecks [5, 6, 7].
4
Kimi K2.5 T ECHNICAL R EPORT
Figure 3: An agent swarm has a trainable orchestrator that dynamically creates specialized frozen subagents and
decomposes complex tasks into parallelizable subtasks for efficient distributed execution.
The limited capacity of a single agent working through each step one by one can lead to the exhaustion of practical
reasoning depth and tool-call budgets, ultimately hindering the system’s ability to handle more complex scenarios.
To address this, we introduce Agent Swarm and Parallel Agent Reinforcement Learning (PARL). Instead of exe-
cuting a task as a reasoning chain or relying on pre-specified parallelization heuristics, K2.5 initiates an Agent Swarm
through dynamic task decomposition, subagent instantiation, and parallel subtask scheduling. Importantly, parallelism
is not presumed to be inherently advantageous; decisions regarding whether, when, and how to parallelize are explic-
itly learned through environmental feedback and RL-driven exploration. As shown in Figure 4, the progression of
performance demonstrates this adaptive capability, with the cumulative reward increasing smoothly as the orchestrator
optimizes its parallelization strategy throughout training.
Architecture and Learning Setup The PARL framework adopts a decoupled architecture comprising a trainable
orchestrator and frozen subagents instantiated from fixed intermediate policy checkpoints. This design deliberately
avoids end-to-end co-optimization to circumvent two fundamental challenges: credit assignment ambiguity and train-
ing instability. In this multi-agent setting, outcome-based rewards are inherently sparse and noisy; a correct final
answer does not guarantee flawless subagent execution, just as a failure does not imply universal subagent error. By
freezing the subagents and treating their outputs as environmental observations rather than differentiable decision
points, we disentangle high-level coordination logic from low-level execution proficiency, leading to more robust con-
vergence. To improve efficiency, we first train the orchestrator using small-size subagents before transitioning to larger
models. Our RL framework also supports dynamically adjusting the inference instance ratios between subagents and
the orchestrator, thereby maximizing the resource usage across the cluster.
PARL Reward Training a reliable parallel orchestrator is challenging due to the delayed, sparse, and non-stationary
feedback inherent in independent subagent execution. To address this, we define the PARL reward as:
rPARL (x, y) = λ1 · rparallel + λ2 · rfinish + rperf (x, y) .
| {z } | {z } | {z }
instantiation reward sub-agent finish rate task-level outcome
The performance reward rperf evaluates the overall success and quality of the solution y for a given task x. This
is augmented by two auxiliary rewards, each addressing a distinct challenge in learning parallel orchestration. The
reward rparallel is introduced to mitigate serial collapse—a local optimum where the orchestrator defaults to single-
agent execution. By incentivizing subagent instantiation, this term encourages the exploration of concurrent scheduling
spaces. The rfinish reward focuses on the successful completion of assigned subtasks. It is used to prevent spurious
parallelism, a reward-hacking behavior in which the orchestrator increases parallel metrics dramatically by spawning
5
Kimi K2.5 T ECHNICAL R EPORT
Figure 4: In our parallel-agent reinforcement learning environment, the training accuracy increases smoothly as train-
ing progresses. At the same time, the level of parallelism during training also gradually increases.
many subagents without meaningful task decomposition. By rewarding completed subtasks, rfinish enforces feasibility
and guides the policy toward valid and effective decompositions.
To ensure the final policy optimizes for the primary objective, the hyperparameters λ1 and λ2 are annealed to zero over
the course of training.
Critical Steps as Resource Constraint To measure computational time cost in a parallel-agent setting, we define
critical steps by analogy to the critical path in a computation graph. We model an episode as a sequence of execution
stages indexed by t = 1, . . . , T . In each stage, the main agent executes an action, which corresponds to either direct tool
(t)
invocation or the instantiation of a group of subagents running in parallel. Let Smain denote the number of steps taken
(t) (t)
by the main agent in stage t (typically Smain = 1), and Ssub,i denote the number of steps taken by the i-th subagent in that
parallel group. The duration of stage t is governed by the longest-running subagent within that cohort. Consequently,
the total critical steps for an episode are defined as
T
(t) (t)
CriticalSteps = ∑ Smain + max Ssub,i .
i
t=1
By constraining training and evaluation using critical steps rather than total steps, the framework explicitly incentivizes
effective parallelization. Excessive subtask creation that does not reduce the maximum execution time of parallel
groups yields little benefit under this metric, while well-balanced task decomposition that shortens the longest parallel
branch directly reduces critical steps. As a result, the orchestrator is encouraged to allocate work across subagents in
a way that minimizes end-to-end latency, rather than merely maximizing concurrency or total work performed.
Prompt Construction for Parallel-agent Capability Induction To incentivize the orchestrator to leverage the ad-
vantages of parallelization, we construct a suite of synthetic prompts designed to stress the limits of sequential agentic
execution. These prompts emphasize either wide search, requiring simultaneous exploration of many independent
information sources, or deep search, requiring multiple reasoning branches with delayed aggregation. We additionally
include tasks inspired by real-world workloads, such as long-context document analysis and large-scale file down-
loading. When executed sequentially, these tasks are difficult to complete within fixed reasoning-step and tool-call
budgets. By construction, they encourage the orchestrator to allocate subtasks in parallel, enabling completion within
fewer critical steps than would be feasible for a single sequential agent. Importantly, the prompts do not explicitly in-
struct the model to parallelize. Instead, they shape the task distribution such that parallel decomposition and scheduling
strategies are naturally favored.
4 Method Overview
4.1 Foundation: Kimi K2 Base Model
The foundation of Kimi K2.5 is Kimi K2 [53], a trillion-parameter mixture-of-experts (MoE) transformer [59] model
pre-trained on 15 trillion high-quality text tokens. Kimi K2 employs the token-efficient MuonClip optimizer [30,
34] with QK-Clip for training stability. The model comprises 1.04 trillion total parameters with 32 billion activated
6
Kimi K2.5 T ECHNICAL R EPORT
Table 3: Overview of training stages: data composition, token volumes, sequence lengths, and trainable components.
Stages ViT Training Joint Pre-training Joint Long-context Mid-training
+ +
Alt text
Text, Knowledge High-quality Text & Multimodal
Data Synthesis Caption
Interleaving Long Text, Long Video
Grounding, OCR, Video
Video, OS Screenshot Reasoning, Long-CoT
Sequence length 4096 4096 32768→262144
Tokens 1T 15T 500B→200B
Training ViT ViT & LLM ViT & LLM
parameters, utilizing 384 experts with 8 activated per token (sparsity of 48). For detailed descriptions of MuonClip,
architecture design, and training infrastructure, we refer to the Kimi K2 technical report [53].
4.2 Model Architecture
The multimodal architecture of Kimi K2.5 consists of three components: a three-dimensional native-resolution vision
encoder (MoonViT-3D), an MLP projector, and the Kimi K2 MoE language model, following the design principles
established in Kimi-VL [54].
MoonViT-3D: Shared Embedding Space for Images and Videos In Kimi-VL, we employ MoonViT to natively
process images at their original resolutions, eliminating the need for complex sub-image splitting and splicing oper-
ations. Initialized from SigLIP-SO-400M [77], MoonViT incorporates the patch packing strategy from NaViT [15],
where single images are divided into patches, flattened, and sequentially concatenated into 1D sequences, thereby
enabling efficient simultaneous training on images at varying resolutions.
To maximize the transfer of image understanding capabilities to video, we introduce MoonViT-3D with a unified
architecture, fully shared parameters, and a consistent embedding space. By generalizing the “patch n’ pack“ philos-
ophy to the temporal dimension, up to four consecutive frames are treated as a spatiotemporal volume: 2D patches
from these frames are jointly flattened and packed into a single 1D sequence, allowing the identical attention mecha-
nism to operate seamlessly across both space and time. While the extra temporal attention improves understanding on
high-speed motions and visual effects, the sharing maximizes knowledge generalization from static images to dynamic
videos, achieving strong video understanding performance (see in Tab. 4) without requiring specialized video mod-
ules or architectural bifurcation. Prior to the MLP projector, lightweight temporal pooling aggregates patches within
each temporal chunk, yielding 4× temporal compression to significantly extend feasible video length. The result is a
unified pipeline where knowledge and ability obtained from image pretraining transfers holistically to videos through
one shared parameter space and feature representation.
4.3 Pre-training Pipeline
As illustrated in Table 3, Kimi K2.5’s pre-training builds upon the Kimi K2 language model checkpoint and processes
approximately 15T tokens across three stages: first, standalone ViT training to establish a robust native-resolution
visual encoder; second, joint pre-training to simultaneously enhance language and multimodal capabilities; and third,
mid-training on high-quality data and long-context activation to refine capabilities and extend context windows.
ViT Training Stage The MoonViT-3D is continual pre-trained from SigLIP [77] on image-text and video-text pairs,
where the text components consist of a variety of targets: image alt texts, synthetic captions of images and videos,
grounding bboxes, and OCR texts. Unlike the implementation in Kimi-VL [54], this continual pre-training does not
include a contrastive loss, but incorporates solely cross-entropy loss Lcaption for caption generation conditioned on
input images and videos. We adopt a two-stage alignment strategy. In the first stage, we update the MoonViT-3D to
align it with Moonlight-16B-A3B [34] via the caption loss, consuming about 1T tokens with very few training FLOPs.
This stage allows MoonViT-3D to primarily understand high-resolution images and videos. A very short second stage
follows, updating only the MLP projector to bridge the ViT with the 1T LLM for smoother joint pre-training.
7
Kimi K2.5 T ECHNICAL R EPORT
Joint Training Stages The joint pre-training stage continues from a near-end Kimi K2 checkpoint over additional
15T vision-text tokens at 4K sequence length. The data recipe extends Kimi K2’s pre-training distribution by in-
troducing unique tokens, adjusting data proportions with increased weight on coding-related content, and controlling
maximum epochs per data source. The third stage performs long-context activation with integrated higher-quality mid-
training data, sequentially extending context length via YaRN [44] interpolation. This yields significant generalization
improvements in long-context text understanding and long video comprehension.
4.4 Post-Training
4.4.1 Supervised Fine-Tuning
Following the SFT pipeline established by Kimi K2 [53], we developed K2.5 by synthesizing high-quality candidate
responses from K2, K2 Thinking and a suite of proprietary in-house expert models. Our data generation strategy
employs specialized pipelines tailored to specific domains, integrating human annotation with advanced prompt en-
gineering and multi-stage verification. This methodology produced a large-scale instruction-tuning dataset featuring
diverse prompts and intricate reasoning trajectories, ultimately training the model to prioritize interactive reasoning
and precise tool-calling for complex, real-world applications.
4.4.2 Reinforcement Learning
Reinforcement learning constitutes a crucial phase of our post-training. To facilitate joint optimization across text and
vision modalities, as well as to enable PARL for agent swarm, we develop a Unified Agentic Reinforcement Learning
Environment (Appendix D) and optimize the RL algorithms. Both text-vision joint RL and PARL are built upon the
algorithms described in this section.
Policy Optimization For each problem x sampled from a dataset D, K responses {y1 , . . . , yK } are generated using
the previous policy πold . We optimize the model πθ with respect to the following objective:
!2
K |y j |
!
1 πθ (yij |x, y0:i
j ) π θ (y i |x, y0:i )
j j
LRL (θ ) = Ex∼D ∑ ∑ Clip , α, β (r(x, y j ) − r̄(x)) − τ log . (1)
N j=1 i=1 πold (yij |x, y0:i
j ) π old (y i |x, y0:i )
j j
j
Here α, β , τ > 0 are hyperparameters, y0:i is the prefix up to the i-th token of the j-th response, N = ∑Ki=1 |yi | is the
total number of generated tokens in a batch, r̄(x) = K1 ∑Kj=1 r(x, y j ) is the mean reward of all generated responses.
This loss function departs from the policy optimization algorithm used in K1.5 [31] by introducing a token-level
clipping mechanism designed to mitigate the off-policy divergence amplified by discrepancies between training and
inference frameworks. The mechanism functions as a simple gradient masking scheme: policy gradients are computed
normally for tokens with log-ratios within the interval [α, β ], while gradients for tokens falling outside this range are
zeroed out. Notably, a key distinction from standard PPO clipping [50] is that our method relies strictly on the log-ratio
to explicitly bound off-policy drift, regardless of the sign of the advantages. This approach aligns with recent strategies
proposed to stabilize large-scale RL training [74, 78]. Empirically, we find this mechanism essential for maintaining
training stability in complex domains requiring long-horizon, multi-step tool-use reasoning. We employ the MuonClip
optimizer [30, 34] to minimize this objective.
Reward Function We apply a rule-based outcome reward for tasks with verifiable solutions, such as reasoning and
agentic tasks. To optimize resource consumption, we also incorporate a budget-control reward aimed at enhancing
token efficiency. For general-purpose tasks, we employ Generative Reward Models (GRMs) that provide granular
evaluations aligned with Kimi’s internal value criteria. In addition, for visual tasks, we design task-specific reward
functions to provide fine-grained supervision. For visual grounding and point localization tasks, we employ an F1-
based reward with soft matching: grounding tasks derive soft matches from Intersection over Union (IoU) and point
tasks derive soft matches from Gaussian-weighted distances under optimal matching. For polygon segmentation tasks,
we rasterize the predicted polygon into a binary mask and compute the segmentation IoU against the ground-truth
mask to assign the reward. For OCR tasks, we adopt normalized edit distance to quantify character-level alignment
between predictions and ground-truth. For counting tasks, rewards are assigned based on the absolute difference
between predictions and ground-truth. Furthermore, we synthesize complex visual puzzle problems and utilize an
LLM verifier (Kimi K2) to provide feedback.
Generative Reward Models Kimi K2 leverages a self-critique rubric reward for open-ended generation [53], and
K2.5 extends this line of work by systematically deploying Generative Reward Models (GRMs) across a broad range
8
Kimi K2.5 T ECHNICAL R EPORT
of agentic behaviors and multimodal trajectories. Rather than limiting reward modeling to conversational outputs,
we apply GRMs on top of verified reward signals in diverse environments, including chat assistants, coding agents,
search agents, and artifact-generating agents. Notably, GRMs function not as binary adjudicators, but as fine-grained
evaluators aligned with Kimi’s values that are critical to user experiences, such as helpfulness, response readiness,
contextual relevance, appropriate level of detail, aesthetic quality of generated artifacts, and strict instruction following.
This design allows the reward signal to capture nuanced preference gradients that are difficult to encode with purely
rule-based or task-specific verifiers. To mitigate reward hacking and overfitting to a single preference signal, we
employ multiple alternative GRM rubrics tailored to different task contexts.
Token Efficient Reinforcement Learning Token efficiency is central to LLMs with test-time scaling. While test-
time scaling inherently trades computation for reasoning quality, practical gains require algorithmic innovations that
actively navigate this trade-off. Our previous findings indicate that imposing a problem-dependent budget effectively
constrains inference-time compute, incentivizing the model to generate more concise chain of thought reasoning pat-
terns without unnecessary token expansion [31, 53]. However, we also observe a length-overfitting phenomenon:
models trained under rigid budget constraints often fail to generalize to higher compute scales. Consequently, they
cannot effectively leverage additional inference-time tokens to solve complex problems, instead defaulting to truncated
reasoning patterns.
To this end, we propose Toggle, a training heuristic that alternates between inference-time scaling and budget-
constrained optimization: for learning iteration t, the reward function is defined by
r(x, y) · I K1 ∑Ki=1 r(x, yi ) < λ or |yi | ≤ budget(x)
if ⌊t/m⌋ (mod 2) = 0 (Phase0)
r̃(x, y) = .
r(x, y) if ⌊t/m⌋ (mod 2) = 1 (Phase1)
where λ and m are hyper-parameters of the algorithm and K is the number of rollouts per problem. Specifically, the
algorithm alternates between two optimization phases every m iterations:
• Phase0 (budget limited phase): The model is trained to solve the problem within a task-dependent token budget.
To prevent a premature sacrifice of quality for efficiency, this constraint is conditionally applied: it is only enforced
when the model’s mean accuracy for a given problem exceeds the threshold λ .
• Phase1 (standard scaling phase): The model generates responses up to the maximum token limit, encouraging
the model to leverage computation for better inference-time scaling.
The problem-dependent budget is estimated from the ρ-th percentile of token lengths among the subset of correct
responses:
budget(x) = Percentile ({|y j | | r(x, yi ) = 1, i = 1, . . . , K}, ρ) . (2)
This budget is estimated once at the beginning of training and remains fixed thereafter. Notably, Toggle functions
as a stochastic alternating optimization for a bi-objective problem. It is specifically designed to reconcile reasoning
capabilities with computational efficiency.
We evaluate the effectiveness of Toggle on K2 Thinking [1]. As shown in Figure 5, we observe a consistent reduction in
output length across nearly all benchmarks. On average, Toggle decreases output tokens by 25∼30% with a negligible
impact on performance. We also observe that redundant patterns in the chain-of-thought, such as repeated verifications
and mechanical calculations, decrease substantially. Furthermore, Toggle shows strong domain generalization. For
example, when trained exclusively on mathematics and programming tasks, the model still achieves consistent token
reductions on GPQA and MMLU-Pro with only marginal degradation in performance (Figure 5).
4.5 Training Infrastructure
Kimi K2.5 inherits the training infrastructure from Kimi K2 [53] with minimal modifications. For multimodal training,
we propose Decoupled Encoder Process, where the vision encoder is incorporated into the existing pipeline with
negligible additional overhead.
4.5.1 Decoupled Encoder Process (DEP)
In a typical multimodal training paradigm utilizing Pipeline Parallelism (PP), the vision encoder and text embedding
are co-located in the first stage of the pipeline (Stage-0). However, due to the inherent variations of multimodal
input size (e.g., image counts and resolutions), Stage-0 suffers from drastic fluctuations in both computational load
and memory usage. This forces existing solutions to adopt custom PP configurations for vision-language models
— for instance, [54] manually adjusts the number of text decoder layers in Stage-0 to reserve memory. While this
9
Kimi K2.5 T ECHNICAL R EPORT
Figure 5: Comparison of model performance and token usage for Kimi K2 Thinking following token-efficient RL.
compromise alleviates memory pressure, it does not fundamentally resolve the load imbalance caused by multimodal
input sizes. More critically, it precludes the direct reuse of parallel strategies that have been highly optimized for
text-only training.
Leveraging the unique topological position of the visual encoder within the computation graph — specifically, its role
as the start of the forward pass and the end of the backward pass — our training uses Decoupled Encoder Process
(DEP), which is composed of three stages in each training step:
• Balanced Vision Forward: We first execute the forward pass for all visual data in the global batch. Because the
vision encoder is small, we replicate it on all GPUs regardless of other parallelism strategies. During this phase,
the forward computational workload is evenly distributed across all GPUs based on load metrics (e.g., image or
patch counts). This eliminates load-imbalance caused by PP and visual token counts. To minimize peak memory
usage, we discard all intermediate activations, retaining only the final output activations. The results are gathered
back to PP Stage-0;
• Backbone Training: This phase performs the forward and backward passes for the main transformer backbone.
By discarding intermediate activations in the preceding phase, we can now fully leverage any efficient parallel
strategies validated in pure text training. After this phase, gradients are accumulated at the visual encoder output;
• Vision Recomputation & Backward: We re-compute the vision encoder forward pass, followed by a backward
pass to compute gradients for parameters in the vision encoder;
DEP not only achieves load-balance, but also decouples the optimization strategy of the vision encoder and the main
backbone. K2.5 seamlessly inherits the parallel strategy of K2, achieving a multimodal training efficiency of 90% rel-
ative to text-only training. We note a concurrent work, LongCat-Flash-Omni [55], shares a similar design philosophy.
5 Evaluations
5.1 Main Results
5.1.1 Evaluation Settings
Benchmarks We evaluate Kimi K2.5 on a comprehensive benchmark suite spanning text-based reasoning, compet-
itive and agentic coding, multimodal understanding (image and video), autonomous agentic execution, and computer
use. Our benchmark taxonomy is organized along the following capability axes:
• Reasoning & General: Humanity’s Last Exam (HLE) [46], AIME 2025 [4], HMMT 2025 (Feb) [58], IMO-
AnswerBench [37], GPQA-Diamond [47], MMLU-Pro [64], SimpleQA Verified [22], AdvancedIF [23], and
LongBench v2 [9].
10
Kimi K2.5 T ECHNICAL R EPORT
• Coding: SWE-Bench Verified [29], SWE-Bench Pro (public) [16], SWE-Bench Multilingual [29], Terminal
Bench 2.0 [39], PaperBench (CodeDev) [52], CyberGym [66], SciCode [56], OJBench (cpp) [65], and Live-
CodeBench (v6) [28].
• Agentic Capabilities: BrowseComp [68], WideSearch [69],DeepSearchQA [60], FinSearchComp (T2&T3) [26],
Seal-0 [45], GDPVal [43].
• Image Understanding: (math & reasoning) MMMU-Pro [75], MMMU (val) [76], CharXiv (RQ) [67], Math-
Vision [61] and MathVista (mini) [36]; (vision knowledge) SimpleVQA [13] and WorldVQA 2 ; (perception)
ZeroBench (w/ and w/o tools) [48], BabyVision [12], BLINK [18] and MMVP [57]; (OCR & document) OCR-
Bench [35], OmniDocBench 1.5 [42] and InfoVQA [38].
• Video Understanding: VideoMMMU [25], MMVU [79], MotionBench [24], Video-MME [17] (with subtitles),
LongVideoBench [70], and LVBench [62].
• Computer Use: OSWorld-Verified [72, 73], and WebArena [80].
Baselines We benchmark against state-of-the-art proprietary and open-source models. For proprietary models, we
compare against Claude Opus 4.5 (with extended thinking) [6], GPT-5.2 (with xhigh reasoning effort) [41], and Gemini
3 Pro (with high reasoning-level) [20]. For open-source models, we include DeepSeek-V3.2 (with thinking mode
enabled) [14] for text benchmarks, while vision benchmarks report Qwen3-VL-235B-A22B-Thinking [8] instead.
Evaluation Configurations Unless otherwise specified, all Kimi K2.5 evaluations use temperature = 1.0, top-p =
0.95, and a context length of 256k tokens. Benchmarks without publicly available scores were re-evaluated under
identical conditions and marked with an asterisk (*). The full evaluation settings can be found in appendix E.
5.1.2 Evaluation Results
Comprehensive results comparing Kimi K2.5 against proprietary and open-source baselines are presented in Table 4.
We highlight key observations across core capability domains:
Reasoning and General Kimi K2.5 achieves competitive performance with top-tier proprietary models on rigorous
STEM benchmarks. On Math tasks, AIME 2025, K2.5 scores 96.1%, approaching GPT-5.2’s perfect score while out-
performing Claude Opus 4.5 (92.8%) and Gemini 3 Pro (95.0%). This high-level performance extends to the HMMT
2025 (95.4%) and IMO-AnswerBench (81.8%), demonstrating K2.5’s superior reasoning depth. Kimi K2.5 also ex-
hibits remarkable knowledge and scientific reasoning capabilities, scoring 36.9% on SimpleQA Verified, 87.1% on
MMLU-Pro and 87.6% on GPQA. Notably, on HLE without the use of tools, K2.5 achieves an HLE-Full score of
30.1%, with component-wise scores of 31.5% on text subset and 21.3% on image subset. When tool-use is enabled,
K2.5’s HLE-Full score rises to 50.2%, with 51.8% (text) and 39.8% (image), significantly outperforming Gemini 3
Pro (45.8%) and GPT-5.2 (45.5%). In addition to reasoning and knowledge, K2.5 shows strong instruction-following
performance (75.6% on AdvancedIF) and competitive long-context abilities, achieving 61.0% on LongBench v2 com-
pared to both proprietary and open-source models.
Complex Coding and Software Engineering Kimi K2.5 exhibits strong software engineering capabilities, espe-
cially on realistic coding and maintenance tasks. It achieves 76.8% on SWE-Bench Verified and 73.0% on SWE-
Bench Multilingual, outperforming Gemini 3 Pro while remaining competitive with Claude Opus 4.5 and GPT-5.2.
On LiveCodeBench v6, Kimi K2.5 reaches 85.0%, surpassing DeepSeek-V3.2 (83.3%) and Claude Opus 4.5 (82.2%),
highlighting its robustness on live, continuously updated coding challenges. On TerminalBench 2.0, PaperBench,
and SciCode, it scores 50.8%, 63.5%, and 48.7% respectively, demonstrating stable competition-level performance
in automated software engineering and problem solving across diverse domains. In addition, K2.5 attains a score of
41.3 on CyberGym, on the task of finding previously discovered vulnerabilities in real open-source software projects
given only a high-level description of the weakness, further underscoring its effectiveness in security-oriented software
analysis.
Agentic Capabilities Kimi K2.5 establishes new state-of-the-art performance on complex agentic search and brows-
ing tasks. On BrowseComp, K2.5 achieves 60.6% without context management techniques, 74.9% with Discard-all
context management [14] — substantially outperforming GPT-5.2’s reported 65.8%, Claude Opus 4.5 (37.0%) and
Gemini 3 Pro (37.8%). Similarly, WideSearch reaches 72.7% on item-f1. On DeepSearchQA (77.1%), FinSearch-
CompT2&T3 (67.8%) and Seal-0 (57.4%), K2.5 leads all evaluated models, demonstrating superior capacity for agen-
tic deep research, information synthesis, and multi-step tool orchestration.
2 https://github.com/MoonshotAI/WorldVQA
11
Kimi K2.5 T ECHNICAL R EPORT
Table 4: Performance comparison of Kimi K2.5 against open-source and proprietary models. Bold denotes the global
SOTA; Data points marked with * are taken from our internal evaluations. † refers to their scores of text-only subset.
Proprietary Open Source
Benchmark Kimi K2.5 Claude Opus GPT-5.2 Gemini 3 Pro DeepSeek- Qwen3-VL-
4.5 (xhigh) V3.2 235B-A22B
Reasoning & General
HLE-Full 30.1 30.8 34.5 37.5 25.1† -
HLE-Full w/ tools 50.2 43.2 45.5 45.8 40.8† -
AIME 2025 96.1 92.8 100 95.0 93.1 -
HMMT 2025 (Feb) 95.4 92.9* 99.4 97.3* 92.5 -
IMO-AnswerBench 81.8 78.5* 86.3 83.1* 78.3 -
GPQA-Diamond 87.6 87.0 92.4 91.9 82.4 -
MMLU-Pro 87.1 89.3* 86.7* 90.1 85.0 -
SimpleQA Verified 36.9 44.1 38.9 72.1 27.5 -
AdvancedIF 75.6 63.1 81.1 74.7 58.8 -
LongBench v2 61.0 64.4* 54.5* 68.2* 59.8* -
Coding
SWE-Bench Verified 76.8 80.9 80.0 76.2 73.1 -
SWE-Bench Pro (public) 50.7 55.4* 55.6 - - -
SWE-Bench Multilingual 73.0 77.5 72.0 65.0 70.2 -
Terminal Bench 2.0 50.8 59.3 54.0 54.2 46.4 -
PaperBench (CodeDev) 63.5 72.9* 63.7* - 47.1 -
CyberGym 41.3 50.6 - 39.9* 17.3* -
SciCode 48.7 49.5 52.1 56.1 38.9 -
OJBench (cpp) 57.4 54.6* - 68.5* 54.7* -
LiveCodeBench (v6) 85.0 82.2* - 87.4* 83.3 -
Agentic
BrowseComp 60.6 37.0 37.8 51.4 -
65.8
BrowseComp (w/ ctx manage) 74.9 57.8 59.2 67.6 -
BrowseComp (Agent Swarm) 78.4 - - - - -
WideSearch 72.7 76.2* - 57.0 32.5* -
WideSearch (Agent Swarm) 79.0 - - - - -
DeepSearchQA 77.1 76.1* 71.3* 63.2* 60.9* -
FinSearchCompT2&T3 67.8 66.2* - 49.9 59.1* -
Seal-0 57.4 47.7* 45.0 45.5* 49.5* -
GDPVal-AA 41.0 45.0 48.0 35.0 34.0 -
Image
MMMU-Pro 78.5 74.0 79.5* 81.0 - 69.3
MMMU (val) 84.3 80.7 86.7* 87.5* - 80.6
CharXiv (RQ) 77.5 67.2* 82.1 81.4 - 66.1
MathVision 84.2 77.1* 83.0 86.1* - 74.6
MathVista (mini) 90.1 80.2* 82.8* 89.8* - 85.8
SimpleVQA 71.2 69.7* 55.8* 69.7* - 56.8*
WorldVQA 46.3 36.8 28.0 47.4 - 23.5
ZeroBench 9 3* 9* 8* - 4*
ZeroBench w/ tools 11 9* 7* 12* - 3*
BabyVision 36.5 14.2 34.4 49.7 - 22.2
BLINK 78.9 68.8* - 78.7* - 68.9
MMVP 87.0 80.0* 83.0* 90.0* - 84.3
OmniDocBench 1.5 88.8 87.7* 85.7 88.5 - 82.0*
OCRBench 92.3 86.5* 80.7* 90.3* - 87.5
InfoVQA (test) 92.6 76.9* 84* 57.2* - 89.5
Video
VideoMMMU 86.6 84.4* 85.9 87.6 - 80.0
MMVU 80.4 77.3* 80.8* 77.5* - 71.1
MotionBench 70.4 60.3* 64.8* 70.3 - -
Video-MME 87.4 77.6* 86.0* 88.4* - 79.0
LongVideoBench 79.8 67.2* 76.5* 77.7* - 65.6*
LVBench 75.9 57.3 - 73.5* - 63.6
Computer Use
OSWorld-Verified 63.3 66.3 8.6* 20.7* - 38.1
WebArena 58.9 63.4* - - - 26.4*
12
Kimi K2.5 T ECHNICAL R EPORT
Table 5: Performance and token efficiency of some reasoning models. Average output token counts (in thousands) are
shown in parentheses.
Benchmark Kimi K2.5 Kimi K2 Gemini-3.0 DeepSeek-V3.2
Thinking Pro Thinking
AIME 2025 96.1 (25k) 94.5 (30k) 95.0 (15k) 93.1 (16k)
HMMT Feb 2025 95.4 (27k) 89.4 (35k) 97.3 (16k) 92.5 (19k)
HMMT Nov 2025 91.1 (24k) 89.2 (32k) 94.5 (15k) 90.2 (18k)
IMO-AnswerBench 81.8 (36k) 78.6 (37k) 83.1 (18k) 78.3 (27k)
LiveCodeBench 85.0 (18k) 82.6 (25k) 87.4 (13k) 83.3 (16k)
GPQA Diamond 87.6 (14k) 84.5 (13k) 91.9 (8k) 82.4 (7k)
HLE-Text 31.5 (24k) 23.9 (29k) 38.4 (13k) 25.1 (21k)
Vision Reasoning, Knowledge and Perception Kimi K2.5 demonstrates strong visual reasoning and world knowl-
edge capabilities. It scores 78.5% on MMMU-Pro, spanning multi-disciplinary multimodal tasks. For world knowl-
edge question answering, K2.5 achieves 71.2% on SimpleVQA and 46.3% on WorldVQA. For visual reasoning, it
achieves 84.2% on MathVision, 90.1% on MathVista (mini), and 36.5% on BabyVision. For OCR and document
understanding, K2.5 delivers outstanding results with 77.5% on CharXiv (RQ), 92.3% on OCRBench, 88.8% on Om-
niDocBench 1.5, and 92.6% on InfoVQA (test). On the challenging ZeroBench, Kimi K2.5 achieves 9% and 11% with
tool augmentation, substantially ahead of competing models. On basic visual perception benchmarks BLINK (78.9%)
and MMVP (87.0%), we also observe competitive performance of Kimi K2.5, demonstrating its robust real-world
visual perceptions.
Video Understanding Kimi K2.5 achieves state-of-the-art performance across diverse video understanding tasks. It
attains 86.6% on VideoMMMU and 80.4% on MMVU, rivaling frontier leaderships. With the context-compression
and dense temporal understanding abilities of MoonViT-3D, Kimi K2.5 also establishes new global SOTA records in
long-video comprehension with 75.9% on LVBench and 79.8% on LongVideoBench by feeding over 2,000 frames,
while demonstrating robust dense-motion understanding at 70.4% on the highly-dimensional MotionBench.
Computer-Use Capability Kimi K2.5 demonstrates state-of-the-art computer-use capability on real-world tasks.
On the computer-use benchmark OSWorld-Verified [72, 73], it achieves a 63.3% success rate relying solely on GUI
actions without external tools. This substantially outperforms open-source models such as Qwen3-VL-235B-A22B
(38.1%) and OpenAI’s computer-use agent framework Operator (o3-based) (42.9%), while remaining competitive
with the current leading CUA model, Claude Opus 4.5 (66.3%). On WebArena [80], an established benchmark for
GUI-based web browsing, Kimi K2.5 achieves a 58.9% success rate, surpassing OpenAI’s Operator (58.1%) and
approaching the performance of Claude Opus 4.5 (63.4%).
5.2 Agent Swarm Results
Benchmarks To rigorously evaluate the effectiveness of the agent swarm framework, we select three representative
benchmarks that collectively cover deep reasoning, large-scale retrieval, and real-world complexity:
• BrowseComp: A challenging deep-research benchmark that requires multi-step reasoning and complex informa-
tion synthesis.
• WideSearch: A benchmark designed to evaluate the ability to perform broad, multi-step information seeking and
reasoning across diverse sources.
• In-house Swarm Bench: An internally developed Swarm benchmark, designed to evaluate the agent swarm per-
formance under real-world, high-complexity conditions. It covers four domains: WildSearch (unconstrained, real-
world information retrieval over the open web), Batch Download (large-scale acquisition of diverse resources),
WideRead (large-scale document comprehension involving more than 100 input documents), and Long-Form Writ-
ing (coherent generation of extensive content exceeding 100k words). This benchmark incorporates extreme-scale
scenarios that stress-test the orchestration, scalability, and coordination capabilities of agent-based systems.
Performance Table 6 presents the performance of Kimi K2.5 Agent Swarm against single-agent configurations and
proprietary baselines. The results demonstrate substantial performance improvements from multi-agent orchestration.
On BrowseComp, Agent Swarm achieves 78.4%, representing a 17.8% absolute gain over the single-agent K2.5
13
Kimi K2.5 T ECHNICAL R EPORT
Table 6: Performance comparison of Kimi K2.5 Agent Swarm against single-agent and proprietary baselines on agentic
search benchmarks. Bold denotes the best result per benchmark.
Benchmark K2.5 Agent Swarm Kimi K2.5 Claude Opus 4.5 GPT-5.2 GPT-5.2 Pro
BrowseComp 78.4 60.6 37.0 65.8 77.9
WideSearch 79.0 72.7 76.2 - -
In-house Swarm Bench 58.3 41.6 45.8 - -
Figure 6: The word cloud visualizes heterogeneous Figure 7: Comparison of Kimi K2.5 performance un-
K2.5-based sub-agents dynamically instantiated by the der Agent Swarm and Discard-all context management
Orchestrator across tests. in BrowseComp.
(60.6%) and surpassing even GPT-5.2 Pro (77.9%). Similarly, WideSearch sees a 6.3% improvement (72.7% →
79.0%) on Item-F1, enabling K2.5 Agent Swarm to outperform Claude Opus 4.5 (76.2%) and establish a new state-
of-the-art. The gains are most pronounced on In-house Swarm bench (16.7%), where tasks are explicitly designed
to reward parallel decomposition. These consistent improvements across benchmarks validate that Agent Swarm
effectively converts computational parallelism into qualitative capability gains, particularly for problems requiring
broad exploration, multi-source verification, or simultaneous handling of independent sub-tasks.
Execution Time Savings via Parallelism Beyond improved task performance, Agent Swarm achieves substantial
wall-clock time reductions through parallel subagent execution. On the WideSearch benchmark, it reduces the exe-
cution time required to reach target performance by 3× ∼ 4.5× compared to a single-agent baseline. As shown in
Figure 8, this efficiency gain scales with task complexity: as the target Item-F1 increases from 30% to 70%, the single
agent’s execution time grows from approximately 1.8× to over 7.0× the baseline, whereas Agent Swarm maintains
near-constant low latency in the range of 0.6× ∼ 1.6×. These results indicate that Agent Swarm effectively transforms
sequential tool invocations into parallel operations, preventing the linear growth in completion time typically observed
as task difficulty increases.
Dynamic Subagent Creation and Scheduling Within an agent swarm, subagents are dynamically instantiated
rather than pre-defined. Through PARL, the orchestrator learns adaptive policies to create and schedule self-hosted
subagents in response to evolving task structures and problem states. Unlike static decomposition approaches, this
learned policy enables the Orchestrator to reason about the requisite number, timing, and specialization of subagents
based on query. Consequently, a heterogeneous agent group emerges organically from this adaptive allocation strategy
(Figure 6).
Agent Swarm as Proactive Context Management Beyond better performance and runtime acceleration, an agent
swarm is a kind of proactive and intelligent context management enabled by multi-agent architecture [5]. This ap-
proach differs from test-time context truncation strategies such as Hide-Tool-Result [2], Summary [71], or Discard-all
[14], which react to context overflow by compressing or discarding accumulated histories. While effective at reducing
token usage, these methods are inherently reactive and often sacrifice structural information or intermediate reasoning.
In contrast, Agent Swarm enables proactive context control through explicit orchestration. Long-horizon tasks are
decomposed into parallel, semantically isolated subtasks, each executed by a specialized subagent with a bounded
local context. Crucially, these subagents maintain independent working memories and perform local reasoning without
directly mutating or contaminating the global context of the central orchestrator. Only task-relevant outputs—rather
than full interaction traces—are selectively routed back to the orchestrator. This design induces context sharding
14
Kimi K2.5 T ECHNICAL R EPORT
Execution Time to Achieve a Target Item-F1
8.0x
Agent Swarm
7.0x Single Agent
6.0x
Execution Time
5.0x
save ×4.5
4.0x
3.0x save ×3.7
2.0x save ×3.2
save ×3.0
save ×3.0
1.0x
0x 30.0% 35.0% 40.0% 45.0% 50.0% 55.0% 60.0% 65.0% 70.0%
Target Item-F1
Figure 8: Agent Swarm achieves 3×–4.5× faster execution time compared to single-agent baselines as target Item-F1
increases from 30% to 70% in WideSearch testing.
rather than context truncation, allowing the system to scale effective context length along an additional architectural
dimension while preserving modularity, information locality, and reasoning integrity.
As shown in Figure 7, this proactive strategy outperforms Discard-all in both efficiency and accuracy on BrowseComp.
By preserving task-level coherence at the orchestrator level while keeping subagent contexts tightly bounded, Agent
Swarm enables parallel execution with selective context persistence, retaining only high-level coordination signals
or essential intermediate results. Consequently, Agent Swarm operates as an active, structured context manager,
achieving higher accuracy with substantially fewer critical steps than uniform context truncation.
6 Conclusions
Kimi K2.5 shows that scalable and general agentic intelligence can be achieved through joint optimization of text and
vision together with parallel agent execution. By unifying language and vision across pre-training and reinforcement
learning, the model achieves strong cross-modal alignment and visual–text reasoning. Agent Swarm enables con-
current execution of heterogeneous sub-tasks, reducing inference latency while improving performance on complex
agentic workloads. Grounded in vision–text intelligence and agent swarms, Kimi K2.5 demonstrates strong perfor-
mance on benchmarks and real-world tasks. By open-sourcing the post-trained checkpoints, we aim to support the
open-source community in building scalable and general-purpose agentic systems and to accelerate progress toward
General Agentic Intelligence.
15
Kimi K2.5 T ECHNICAL R EPORT
References
[1] Moonshot AI. Introducing Kimi K2 Thinking. 2025. URL: https://moonshotai.github.io/Kimi-
K2/thinking.html.
[2] Moonshot AI. Kimi-Researcher End-to-End RL Training for Emerging Agentic Capabilities. 2025. URL:
https://moonshotai.github.io/Kimi-Researcher/.
[3] Amazon Web Services. Amazon Simple Storage Service (Amazon S3). Web. Available at: https://aws.
amazon.com/s3/. 2023. URL: https://aws.amazon.com/s3/ (visited on 12/15/2023).
[4] Mathematical Association of America. 2025 American Invitational Mathematics Examination I. Held on Febru-
ary 6, 2025. 2025. URL: https://artofproblemsolving.com/wiki/index.php/2025_AIME_
I.
[5] Anthropic. Building multi-agent systems: when and how to use them. 2026. URL: https://claude.com/
blog/building-multi-agent-systems-when-and-how-to-use-them.
[6] Anthropic. Claude Opus 4.5 System Card. 2025. URL: https : / / www - cdn . anthropic . com /
bf10f64990cfda0ba858290be7b8cc6317685f47.pdf.
[7] Anthropic. How we built our multi-agent research system. 2025. URL: https://www.anthropic.com/
engineering/multi-agent-research-system.
[8] Shuai Bai et al. Qwen3-VL Technical Report. 2025. arXiv: 2511 . 21631 [cs.CV]. URL: https : / /
arxiv.org/abs/2511.21631.
[9] Yushi Bai et al. LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Mul-
titasks. 2025. arXiv: 2412.15204 [cs.CL]. URL: https://arxiv.org/abs/2412.15204.
[10] Greg Brockman et al. OpenAI Gym. 2016. arXiv: 1606.01540 [cs.LG]. URL: https://arxiv.org/
abs/1606.01540.
[11] Tom B. Brown et al. Language Models are Few-Shot Learners. 2020. arXiv: 2005.14165 [cs.CL]. URL:
https://arxiv.org/abs/2005.14165.
[12] Liang Chen et al. BabyVision: Visual Reasoning Beyond Language. 2026. arXiv: 2601.06521 [cs.CV].
URL: https://arxiv.org/abs/2601.06521.
[13] Xianfu Cheng et al. SimpleVQA: Multimodal Factuality Evaluation for Multimodal Large Language Models.
2025. arXiv: 2502.13059 [cs.CL]. URL: https://arxiv.org/abs/2502.13059.
[14] DeepSeek-AI et al. DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models. 2025. arXiv: 2512.
02556 [cs.CL]. URL: https://arxiv.org/abs/2512.02556.
[15] Mostafa Dehghani et al. Patch n’ Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution. 2023.
arXiv: 2307.06304 [cs.CV]. URL: https://arxiv.org/abs/2307.06304.
[16] Xiang Deng et al. “SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?” In:
arXiv preprint arXiv:2509.16941 (2025).
[17] Chaoyou Fu et al. Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in
Video Analysis. 2025. arXiv: 2405.21075 [cs.CV]. URL: https://arxiv.org/abs/2405.21075.
[18] Xingyu Fu et al. BLINK: Multimodal Large Language Models Can See but Not Perceive. 2024. arXiv: 2404.
12390 [cs.CV]. URL: https://arxiv.org/abs/2404.12390.
[19] Samir Yitzhak Gadre et al. “Datacomp: In search of the next generation of multimodal datasets”. In: Advances
in Neural Information Processing Systems 36 (2024).
[20] Google. Gemini 3 Pro. 2025. URL: https://deepmind.google/models/gemini/pro/.
[21] Dong Guo et al. Seed1.5-VL Technical Report. 2025. arXiv: 2505 . 07062 [cs.CV]. URL: https : / /
arxiv.org/abs/2505.07062.
[22] Lukas Haas et al. SimpleQA Verified: A Reliable Factuality Benchmark to Measure Parametric Knowledge.
2025. arXiv: 2509.07968 [cs.CL]. URL: https://arxiv.org/abs/2509.07968.
[23] Yun He et al. AdvancedIF: Rubric-Based Benchmarking and Reinforcement Learning for Advancing LLM In-
struction Following. 2025. arXiv: 2511.10507 [cs.CL]. URL: https://arxiv.org/abs/2511.
10507.
[24] Wenyi Hong et al. MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for
Vision Language Models. 2025. arXiv: 2501 . 02955 [cs.CV]. URL: https : / / arxiv . org / abs /
2501.02955.
[25] Kairui Hu et al. Video-MMMU: Evaluating Knowledge Acquisition from Multi-Discipline Professional Videos.
2025. arXiv: 2501.13826 [cs.CV]. URL: https://arxiv.org/abs/2501.13826.
16
Kimi K2.5 T ECHNICAL R EPORT
[26] Liang Hu et al. FinSearchComp: Towards a Realistic, Expert-Level Evaluation of Financial Search and Rea-
soning. 2025. arXiv: 2509.13160 [cs.CL]. URL: https://arxiv.org/abs/2509.13160.
[27] Yanping Huang et al. GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism. 2019.
arXiv: 1811.06965 [cs.CV]. URL: https://arxiv.org/abs/1811.06965.
[28] Naman Jain et al. “Livecodebench: Holistic and contamination free evaluation of large language models for
code”. In: arXiv preprint arXiv:2403.07974 (2024).
[29] Carlos E Jimenez et al. “Swe-bench: Can language models resolve real-world github issues?” In: arXiv preprint
arXiv:2310.06770 (2023).
[30] Keller Jordan et al. Muon: An optimizer for hidden layers in neural networks. 2024. URL: https : / /
kellerjordan.github.io/posts/muon/.
[31] Kimi Team. “Kimi k1. 5: Scaling reinforcement learning with llms”. In: arXiv preprint arXiv:2501.12599
(2025).
[32] Hugo Laurençon et al. “Obelics: An open web-scale filtered dataset of interleaved image-text documents”. In:
Advances in Neural Information Processing Systems 36 (2024).
[33] Dmitry Lepikhin et al. “Gshard: Scaling giant models with conditional computation and automatic sharding”.
In: arXiv preprint arXiv:2006.16668 (2020).
[34] Jingyuan Liu et al. “Muon is Scalable for LLM Training”. In: arXiv preprint arXiv:2502.16982 (2025).
[35] Yuliang Liu et al. “OCRBench: on the hidden mystery of OCR in large multimodal models”. In: Science China
Information Sciences 67.12 (Dec. 2024). ISSN: 1869-1919. DOI: 10.1007/s11432-024-4235-6. URL:
http://dx.doi.org/10.1007/s11432-024-4235-6.
[36] Pan Lu et al. MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts. 2024.
arXiv: 2310.02255 [cs.CV]. URL: https://arxiv.org/abs/2310.02255.
[37] Thang Luong et al. “Towards Robust Mathematical Reasoning”. In: Proceedings of the 2025 Conference on
Empirical Methods in Natural Language Processing. Ed. by Christos Christodoulopoulos et al. Suzhou, China:
Association for Computational Linguistics, Nov. 2025, pp. 35418–35442. ISBN: 979-8-89176-332-6. DOI: 10.
18653 / v1 / 2025 . emnlp - main . 1794. URL: https : / / aclanthology . org / 2025 . emnlp -
main.1794/.
[38] Minesh Mathew et al. InfographicVQA. 2021. arXiv: 2104.12756 [cs.CV]. URL: https://arxiv.
org/abs/2104.12756.
[39] Mike A Merrill et al. “Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line
Interfaces”. In: arXiv preprint arXiv:2601.11868 (2026).
[40] Deepak Narayanan et al. Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-
LM. 2021. arXiv: 2104.04473 [cs.CL]. URL: https://arxiv.org/abs/2104.04473.
[41] OpenAI. Introducing GPT 5.2. 2025. URL: https://openai.com/index/introducing- gpt- 5-
2/.
[42] Linke Ouyang et al. OmniDocBench: Benchmarking Diverse PDF Document Parsing with Comprehensive An-
notations. 2025. arXiv: 2412.07626 [cs.CV]. URL: https://arxiv.org/abs/2412.07626.
[43] Tejal Patwardhan et al. GDPval: Evaluating AI Model Performance on Real-World Economically Valuable
Tasks. 2025. arXiv: 2510.04374 [cs.LG]. URL: https://arxiv.org/abs/2510.04374.
[44] Bowen Peng et al. “Yarn: Efficient context window extension of large language models”. In: arXiv preprint
arXiv:2309.00071 (2023).
[45] Thinh Pham et al. SealQA: Raising the Bar for Reasoning in Search-Augmented Language Models. Seal-0 is
the main subset of this benchmark. 2025. arXiv: 2506.01062 [cs.CL]. URL: https://arxiv.org/
abs/2506.01062.
[46] Long Phan et al. Humanity’s Last Exam. 2025. arXiv: 2501.14249 [cs.LG]. URL: https://arxiv.
org/abs/2501.14249.
[47] David Rein et al. “Gpqa: A graduate-level google-proof q&a benchmark”. In: First Conference on Language
Modeling. 2024.
[48] Jonathan Roberts et al. ZeroBench: An Impossible Visual Benchmark for Contemporary Large Multimodal Mod-
els. 2025. arXiv: 2502.09696 [cs.CV]. URL: https://arxiv.org/abs/2502.09696.
[49] Christoph Schuhmann et al. “Laion-5b: An open large-scale dataset for training next generation image-text
models”. In: Advances in Neural Information Processing Systems 35 (2022), pp. 25278–25294.
[50] John Schulman et al. “Proximal Policy Optimization Algorithms”. In: arXiv preprint arXiv:1707.06347 (2017).
URL: https://arxiv.org/abs/1707.06347.
17
Kimi K2.5 T ECHNICAL R EPORT
[51] Tianhui Song et al. Towards Pixel-Level VLM Perception via Simple Points Prediction. 2026. arXiv: 2601.
19228 [cs.CV]. URL: https://arxiv.org/abs/2601.19228.
[52] Giulio Starace et al. “PaperBench: Evaluating AI’s Ability to Replicate AI Research”. In: arXiv preprint
arXiv:2504.01848 (2025).
[53] Kimi Team et al. “Kimi k2: Open agentic intelligence”. In: arXiv preprint arXiv:2507.20534 (2025).
[54] Kimi Team et al. “Kimi-vl technical report”. In: arXiv preprint arXiv:2504.07491 (2025).
[55] Meituan LongCat Team et al. “Longcat-flash-omni technical report”. In: arXiv preprint arXiv:2511.00279
(2025).
[56] Minyang Tian et al. “Scicode: A research coding benchmark curated by scientists”. In: Advances in Neural
Information Processing Systems 37 (2024), pp. 30624–30650.
[57] Shengbang Tong et al. Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs. 2024. arXiv:
2401.06209 [cs.CV]. URL: https://arxiv.org/abs/2401.06209.
[58] Harvard-MIT Mathematics Tournament. Harvard-MIT Mathematics Tournament, February 2025. Held on
February 15, 2025. 2025. URL: https://www.hmmt.org/www/archive/282.
[59] Ashish Vaswani et al. “Attention is All you Need”. In: Advances in Neural Information Processing Systems.
Ed. by I. Guyon et al. Vol. 30. Curran Associates, Inc., 2017. URL: https://proceedings.neurips.
cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf.
[60] Nikhita Vedula et al. DeepSearchQA: Bridging the Comprehensiveness Gap for Deep Research Agents.
2025. URL: https : / / storage . googleapis . com / deepmind - media / DeepSearchQA /
DeepSearchQA_benchmark_paper.pdf.
[61] Ke Wang et al. Measuring Multimodal Mathematical Reasoning with MATH-Vision Dataset. 2024. arXiv:
2402.14804 [cs.CV]. URL: https://arxiv.org/abs/2402.14804.
[62] Weihan Wang et al. LVBench: An Extreme Long Video Understanding Benchmark. 2025. arXiv: 2406.08035
[cs.CV]. URL: https://arxiv.org/abs/2406.08035.
[63] Xinyuan Wang et al. OpenCUA: Open Foundations for Computer-Use Agents. 2025. arXiv: 2508 . 09123
[cs.AI]. URL: https://arxiv.org/abs/2508.09123.
[64] Yubo Wang et al. MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Bench-
mark. 2024. arXiv: 2406.01574 [cs.CL]. URL: https://arxiv.org/abs/2406.01574.
[65] Zhexu Wang et al. “OJBench: A Competition Level Code Benchmark For Large Language Models”. In: arXiv
preprint arXiv:2506.16395 (2025).
[66] Zhun Wang et al. “CyberGym: Evaluating AI Agents’ Cybersecurity Capabilities with Real-World Vulnerabili-
ties at Scale”. In: arXiv preprint arXiv:2506.02548 (2025).
[67] Zirui Wang et al. CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs. 2024. arXiv:
2406.18521 [cs.CL]. URL: https://arxiv.org/abs/2406.18521.
[68] Jason Wei et al. BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents. 2025. arXiv: 2504.
12516 [cs.CL]. URL: https://arxiv.org/abs/2504.12516.
[69] Ryan Wong et al. WideSearch: Benchmarking Agentic Broad Info-Seeking. 2025. arXiv: 2508 . 07999
[cs.CL]. URL: https://arxiv.org/abs/2508.07999.
[70] Haoning Wu et al. LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understand-
ing. 2024. arXiv: 2407.15754 [cs.CV]. URL: https://arxiv.org/abs/2407.15754.
[71] Xixi Wu et al. ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization. 2025. arXiv:
2509.13313 [cs.CL]. URL: https://arxiv.org/abs/2509.13313.
[72] Tianbao Xie et al. “Introducing OSWorld-Verified”. In: xlang.ai (July 2025). URL: https://xlang.ai/
blog/osworld-verified.
[73] Tianbao Xie et al. OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Envi-
ronments. 2024. arXiv: 2404.07972 [cs.AI].
[74] Feng Yao et al. Your Efficient RL Framework Secretly Brings You Off-Policy RL Training. Aug. 2025. URL:
https://fengyao.notion.site/off-policy-rl.
[75] Xiang Yue et al. MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark. 2025.
arXiv: 2409.02813 [cs.CL]. URL: https://arxiv.org/abs/2409.02813.
[76] Xiang Yue et al. “MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark
for Expert AGI”. In: Proceedings of CVPR. 2024.
[77] Xiaohua Zhai et al. Sigmoid Loss for Language Image Pre-Training. 2023. arXiv: 2303.15343 [cs.CV].
URL: https://arxiv.org/abs/2303.15343.
18
Kimi K2.5 T ECHNICAL R EPORT
[78] Xin Zhao et al. Small Leak Can Sink a Great Ship–Boost RL Training on MoE with IcePop! Sept. 2025. URL:
https://ringtech.notion.site/icepop.
[79] Yilun Zhao et al. MMVU: Measuring Expert-Level Multi-Discipline Video Understanding. 2025. arXiv: 2501.
12380 [cs.CV]. URL: https://arxiv.org/abs/2501.12380.
[80] Shuyan Zhou et al. “WebArena: A Realistic Web Environment for Building Autonomous Agents”. In: arXiv
preprint arXiv:2307.13854 (2023). URL: https://webarena.dev.
[81] Wanrong Zhu et al. “Multimodal c4: An open, billion-scale corpus of images interleaved with text”. In: Ad-
vances in Neural Information Processing Systems 36 (2024).
19
Kimi K2.5 T ECHNICAL R EPORT
A Contributors
Tongtong Bai Zhuoma Gongque Liang Liu Tongyu Sun Junyan Wu Xiaokun Yuan
Yifan Bai Qizheng Gu Shaowei Liu Flood Sung Rucong Wu Yang Yue
Yiping Bao Xinran Gu Shudong Liu Yunpeng Tai Wenhao Wu Weihao Zeng
S.H. Cai Yicheng Gu Shuran Liu Chuning Tang Yuefeng Wu Dunyuan Zha
Yuan Cao Longyu Guan Tianwei Liu Heyi Tang Yuhao Wu Haobing Zhan
Y. Charles Yuanying Guo Tianyu Liu Xiaojuan Tang Yuxin Wu Dehao Zhang
H.S. Che Xiaoru Hao Weizhou Liu Zhengyang Tang Zijian Wu Hao Zhang
Cheng Chen Weiran He Xiangyan Liu Jiawen Tao Chenjun Xiao Jin Zhang
Guanduo Chen Wenyang He Yangyang Liu Shiyuan Teng Jin Xie Puqi Zhang
Huarong Chen Yunjia He Yanming Liu Chaoran Tian Xiaotong Xie Qiao Zhang
Jia Chen Chao Hong Yibo Liu Pengfei Tian Yuchong Xie Rui Zhang
Jiahao Chen Hao Hu Yuanxin Liu Ao Wang Yifei Xin Xiaobin Zhang
Jianlong Chen Jiaxi Hu Yue Liu Bowen Wang Bowei Xing Y. Zhang
Jun Chen Yangyang Hu Zhengying Liu Chensi Wang Boyu Xu Yadong Zhang
Kefan Chen Zhenxing Hu Zhongnuo Liu Chuang Wang Jianfan Xu Yangkun Zhang
Liang Chen Ke Huang Enzhe Lu Congcong Wang Jing Xu Yichi Zhang
Ruijue Chen Ruiyuan Huang Haoyu Lu Dingkun Wang Jinjing Xu Yizhi Zhang
Xinhao Chen Weixiao Huang Zhiyuan Lu Dinglu Wang L.H. Xu Yongting Zhang
Yanru Chen Zhiqi Huang Junyu Luo Dongliang Wang Lin Xu Yu Zhang
Yanxu Chen Tao Jiang Tongxu Luo Feng Wang Suting Xu Yushun Zhang
Yicun Chen Zhejun Jiang Yashuo Luo Hailong Wang Weixin Xu Yutao Zhang
Yimin Chen Xinyi Jin Long Ma Haiming Wang Xinbo Xu Yutong Zhang
Yingjiang Chen Yu Jing Yingwei Ma Hengzhi Wang Xinran Xu Zheng Zhang
Yuankun Chen Guokun Lai Shaoguang Mao Huaqing Wang Yangchuan Xu Chenguang Zhao
Yujie Chen Aidi Li Yuan Mei Hui Wang Yichang Xu Feifan Zhao
Yutian Chen C. Li Xin Men Jiahao Wang Yuemeng Xu Jinxiang Zhao
Zhirong Chen Cheng Li Fanqing Meng Jinhong Wang Zelai Xu Shuai Zhao
Ziwei Chen Fang Li Zhiyong Meng Jiuzheng Wang Ziyao Xu Xiangyu Zhao
Dazhi Cheng Guanghe Li Yibo Miao Kaixin Wang Junjie Yan Yikai Zhao
Minghan Chu Guanyu Li Minqing Ni Linian Wang Yuzi Yan Zijia Zhao
Jialei Cui Haitao Li Kun Ouyang Qibin Wang Guangyao Yang Huabin Zheng
Jiaqi Deng Haoyang Li Siyuan Pan Shengjie Wang Hao Yang Ruihan Zheng
Muxi Diao Jia Li Bo Pang Shuyi Wang Junwei Yang Shaojie Zheng
Hao Ding Jingwei Li Yuchao Qian Si Wang Kai Yang Tengyang Zheng
Mengfan Dong Junxiong Li Ruoyu Qin Wei Wang Ningyuan Yang Junfeng Zhong
Mengnan Dong Lincan Li Zeyu Qin Xiaochen Wang Ruihan Yang Longguang Zhong
Yuxin Dong Mo Li Jiezhong Qiu Xinyuan Wang Xiaofei Yang Weiming Zhong
Yuhao Dong Weihong Li Bowen Qu Yao Wang Xinlong Yang M. Zhou
Ang’ang Du Wentao Li Zeyu Shang Yejie Wang Ying Yang Runjie Zhou
Chenzhuang Du Xinhang Li Youbo Shao Yipu Wang Yi (弋) Yang Xinyu Zhou
Dikang Du Xinhao Li Tianxiao Shen Yiqin Wang Yi (翌) Yang Zaida Zhou
Lingxiao Du Yang Li Zhennan Shen Yucheng Wang Zhen Yang Jinguo Zhu
Yulun Du Yanhao Li Juanfeng Shi Yuzhi Wang Zhilin Yang Liya Zhu
Yu Fan Yiwei Li Lidong Shi Zhaoji Wang Zonghan Yang Xinhao Zhu
Shengjun Fang Yuxiao Li Shengyuan Shi Zhaowei Wang Haotian Yao Yuxuan Zhu
Qiulin Feng Zhaowei Li Feifan Song Zhengtao Wang Dan Ye Zhen Zhu
Yichen Feng Zheming Li Pengwei Song Zhexu Wang Wenjie Ye Jingze Zhuang
Garimugai Fu Weilong Liao Tianhui Song Zihan Wang Zhuorui Ye Weiyu Zhuang
Kelin Fu Jiawei Lin Xiaoxi Song Zizhe Wang Bohong Yin Ying Zou
Hongcheng Gao Xiaohan Lin Hongjin Su Chu Wei Chengzhen Yu Xinxing Zu
Tong Gao Zhishan Lin Jianlin Su Ming Wei Longhui Yu Kimi K2
Yuyao Ge Zichao Lin Zhaochen Su Chuan Wen Tao Yu† Kimi K2.5
Shangyi Geng Cheng Liu Lin Sui Zichen Wen Tianxiang Yu
Chengyang Gong Chenyu Liu Jinsong Sun Chengjie Wu Enming Yuan
Xiaochen Gong Hongzhang Liu Junyao Sun Haoning Wu Mengjie Yuan
The listing of authors is in alphabetical order based on their last names.
† The University of Hong Kong
20
Kimi K2.5 T ECHNICAL R EPORT
Vision Knowledge Vision General Reasoning OCR
Score Score Score
Vision:Text = 10%:90% Vision:Text = 10%:90% Vision:Text = 10%:90%
Vision:Text = 20%:80% Vision:Text = 20%:80% Vision:Text = 20%:80%
Vision:Text = 50%:50% Vision:Text = 50%:50% Vision:Text = 50%:50%
Steps Steps Steps
Text Knowledge Text General Reasoning Coding
Score Score Score
Vision:Text = 10%:90% Vision:Text = 10%:90% Vision:Text = 10%:90%
Vision:Text = 20%:80% Vision:Text = 20%:80% Vision:Text = 20%:80%
Vision:Text = 50%:50% Vision:Text = 50%:50% Vision:Text = 50%:50%
Steps Steps Steps
Figure 9: Learning curves comparing vision-to-text ratios (10:90, 20:80, 50:50) under fixed vision-text token budget
across vision and language tasks. Early fusion with lower vision ratios tend to yield better results.
B Pre-training
B.1 Joint-Training
We further provide the full training curves for all configurations in Figure 9. Notably, we observe a "dip-and-recover"
pattern in text performance during mid-fusion and late-fusion stages: when vision data is first introduced, text ca-
pability initially degrades before gradually recovering. We attribute this to the modality domain shift—the sudden
introduction of vision tokens disrupts the established linguistic representation space, forcing the model to temporarily
sacrifice text-specific competence for cross-modal alignment.
In contrast, early fusion maintains a healthier and more stable text performance curve throughout training. By co-
optimizing vision and language from the outset, the model naturally evolves unified multimodal representations with-
out the shock of late-stage domain migration. This suggests that early exposure not only prevents the representation
collapse observed in late fusion but also facilitates smoother gradient landscapes for both modalities. Collectively,
these findings reinforce our proposal of native multimodal pre-training: moderate vision ratios combined with early
fusion yield superior convergence properties and more robust bi-modal competence under fixed token budgets.
B.2 Text data
The Kimi K2.5 pre-training text corpus comprises curated, high-quality data spanning four primary domains: Web
Text, Code, Mathematics, and Knowledge. Most data processing pipelines follow the methodologies outlined in Kimi
K2 [53]. For each domain, we performed rigorous correctness and quality validation and designed targeted data
experiments to ensure the curated dataset achieved both high diversity and effectiveness.
Enhanced Code Intelligence We upweighted code-centric data, significantly expanding (1) repository-level code sup-
porting cross-file reasoning and architectural understanding, (2) issues, code reviews and commit histories from the
internet capturing real-world development patterns, and (3) code-related documents retrieved from PDF and webtext
corpora. These efforts strengthen repository-level comprehension for complex coding tasks, improve performance on
agentic coding subtasks such as patch generation and unit test writing, and enhance code-related knowledge capabili-
ties.
21
Kimi K2.5 T ECHNICAL R EPORT
B.3 Vision data
Our multimodal pre-training corpus includes seven categories: caption, interleaving, OCR, knowledge, perception,
video, and agent data. Caption data [49, 19] provides fundamental modality alignment, with strict limits on synthetic
captions to mitigate hallucination. Image-text interleaving data from books, web pages, and tutorials [81, 32] enables
multi-image comprehension and longer context learning. OCR data spans multilingual text, dense layouts, and multi-
page documents. Knowledge data incorporates academic materials processed via layout parsers to develop visual
reasoning capabilities.
Furthermore, we curate a specialized multimodal problem-solving corpus to bolster reasoning within Science, Tech-
nology, Engineering, and Mathematics domains. This data is aggregated through targeted retrieval and web crawling;
for informational content lacking explicit query formats, we employ in-context learning [11] to automatically refor-
mulate raw materials into structured academic problems spanning K-12 to university levels. To bridge the modality
gap between visual layouts and code data, we incorporate extensive image-code paired data. This includes a diverse
array of code formats—such as HTML, React, and SVG, among others—paired with their corresponding rendered
screenshots, enabling the model to align abstract structural logic with concrete visual geometry.
For agentic and temporal understanding, we collect GUI screenshots and action trajectories across desktop, mobile, and
web environments, including human-annotated demonstrations. Video data from diverse sources enables both hour-
long video comprehension and fine-grained spatio-temporal perception. Additionally, we incorporate grounding data
to enhance fine-grained visual localization, including perception annotations (bounding boxes), point-based references.
We also introduce a new contour-level segmentation task [51] for pixel-level perception learning. All data undergoes
rigorous filtering, deduplication, and quality control to ensure high diversity and effectiveness.
C Infra
Kimi K2.5 is trained on NVIDIA H800 GPU clusters with 8×400 Gbps RoCE interconnects across nodes. We em-
ploy a flexible parallelism strategy combining 16-way Pipeline Parallelism (PP) with virtual stages [27, 40], 16-way
Expert Parallelism (EP) [33], and ZeRO-1 Data Parallelism, enabling training on any number of nodes that is a mul-
tiple of 32. EP all-to-all communication is overlapped with computation under interleaved 1F1B scheduling. To
fit activations within GPU memory constraints, we apply selective recomputation for LayerNorm, SwiGLU, and
MLA up-projections, compress insensitive activations to FP8-E4M3, and offload remaining activations to CPU with
overlapped streaming.
C.1 Data Storage and Loading
We employ S3 [3] compatible object storage solutions from cloud providers to house our VLM datasets. To bridge
the gap between data preparation and model training, we retain visual data in its native format and have engineered a
highly efficient and adaptable data loading infrastructure. This infrastructure offers several critical advantages:
• Flexibility: Facilitates dynamic data shuffling, blending, tokenization, loss masking, and sequence packing
throughout the training process, enabling adjustable data ratios as requirements evolve;
• Augmentation: Allows for stochastic augmentation of both visual and textual modalities, while maintaining the
integrity of 2D spatial coordinates and orientation metadata during geometric transformations;
• Determinism: Guarantees fully deterministic training through meticulous management of random seeds and
worker states, ensuring that any training interruption can be resumed seamlessly — the data sequence after re-
sumption remains identical to an uninterrupted run;
• Scalability: Achieves superior data loading throughput via tiered caching mechanisms, robustly scaling to large
distributed clusters while regulating request frequency to object storage within acceptable bounds.
Furthermore, to uphold uniform dataset quality standards, we have built a unified platform overseeing data registration,
visualization, statistical analysis, cross-cloud synchronization, and lifecycle governance.
D Unified Agentic Reinforcement Learning Environment
Environment To support unified Agentic RL, our RL framework features a standardized Gym-like [10] interface
to streamline the implementation of diverse environments. Such design empowers users to implement and customize
22
Kimi K2.5 T ECHNICAL R EPORT
Recursive Call
Single Agent Task
Token-in Inference
Core Agent Loop Engine
Pluggable Components
Token-out Service
Toolset
Rollout Obs Act A Obs
Act
Manager Judge Mismatch Correction
Prompt & Instruction Black-Box Env
Enhancement
White-Box Env
Training
Engine
LLM Gateway Env Pool Service
Figure 10: Overview of our agentic RL framework.
environments with minimal overhead. Our design prioritizes compositional modularity by integrating a suite of plug-
gable components, such as a Toolset module for supporting various tools with sandboxes, a Judge module for multi-
faceted reward signals, and specialized modules for prompt diversification and instruction-following enhancement.
These components can be dynamically composed with core agent loops, offering high flexibility and enhancing model
generalization.
At the execution level, our RL framework treats every agent task as an independent asynchronous coroutine. Each
task can recursively trigger sub-task rollouts, simplifying the implementation of complex multi-agent paradigms such
as Parallel-Agent RL and Agent-as-Judge. As shown in the figure 10, a dedicated Rollout Manager orchestrates
up to 100,000 concurrent agent tasks during the RL process, providing fine-grained control to enable features like
partial rollout [31]. Upon activation, each task acquires an environment instance from a managed pool, equipped with
a sandbox and specialized tools.
Inference Engine Co-design Our framework strictly follows a Token-in-Token-out paradigm. We also record log
probabilities for all inference engine outputs to perform train-inference mismatch correction, ensuring stable RL train-
ing. A co-design of inference engine for RL requirements has allowed us to support these features by custom inference
APIs for RL.
Besides a comprehensive suite of built-in white-box environments, there are also black-box environments that can
only run under standard LLM API protocol, missing the opportunity to use advanced features offered by our custom
API protocol. To facilitate model optimization under black-box environments, we developed LLM Gateway, which is
a proxy service that keeps detailed records of rollout requests and responses under our custom protocol.
Monitoring and debugging It is a challenging task to optimize performance of a highly-parallel asynchronous
execution system, while ensuring correctness. We develop a series of tools for performance monitoring, profiling, data
visualization and data verification. We found these to be instrumental in debugging and ensuring both the efficiency
and correctness of our Agentic RL.
E Evaluation Settings
This section provides comprehensive configuration details and testing protocols for all benchmarks reported in Table 4.
E.1 General Evaluation Protocol
Unless explicitly stated otherwise, all experiments for Kimi-K2.5 adhere to the following hyperparameter configura-
tion:
• Temperature: 1.0
• Top-p: 0.95
• Context Length: 256k tokens
23
Kimi K2.5 T ECHNICAL R EPORT
E.2 Baselines
For baseline models, we report results under their respective high-performance reasoning configurations:
• Claude Opus 4.5: Extended thinking mode
• GPT-5.2: Maximum reasoning effort (xhigh)
• Gemini 3 Pro: High thinking level
• DeepSeek-V3.2: Thinking mode enabled (for text-only benchmarks)
• Qwen3-VL-235B-A22B: Thinking mode (for vision benchmarks only)
For vision and multimodal benchmarks, GPT-5.2-xhigh exhibited an approximate 10% failure rate (i.e., no output
generated despite three retry attempts) during vision evaluations. These failures were treated as incorrect predictions,
meaning that the reported scores may be conservative lower bounds of the model’s true capability.
In addition, because we were unable to consistently access a stable GPT-5.2 API, we skipped some benchmarks with
high evaluation costs, such as WideSearch.
E.3 Text Benchmarks
Reasoning Benchmarks. For high-complexity reasoning benchmarks, including HLE-Full, AIME 2025, HMMT
2025, GPQA-Diamond, and IMO-AnswerBench, we enforce a maximum completion budget of 96k tokens to ensure
sufficient reasoning depth. To reduce variance arising from stochastic reasoning paths, results on AIME 2025 and
HMMT 2025 (Feb) are averaged over 64 independent runs (Avg@64), while GPQA-Diamond is averaged over 8 runs
(Avg@8).
LongBench v2. For a fair comparison, we standardize all input contexts to approximately 128k tokens using the
same truncation strategy as in [9]. We observe that GPT5.2-xhigh frequently produces free-form question–answer
style responses rather than the required multiple-choice format. Therefore, we report results using GPT5.2-high,
which consistently adheres to the expected output format.
E.4 Image and Video Benchmarks
All image and video understanding evaluations utilize the following configuration:
• Maximum Tokens: 64k
• Sampling: Averaged over 3 independent runs (Avg@3)
ZeroBench (w/ tools). Multi-step reasoning evaluations use constrained step-wise generation:
• Max Tokens per Step: 24k
• Maximum Steps: 30
MMMU-Pro. We adhere strictly to the official evaluation protocol: input order is preserved for all modalities, with
images prepended to text sequences as specified in the benchmark guidelines.
Sampling Strategies for Video Benchmarks. For short video benchmarks (VideoMMMU, MMVU & Motion-
Bench), we sample 128 uniform input frames with a maximum spatial resolution at 896; 2048 uniform frames are
sampled for long video benchmarks (Video-MME, LongVideoBench & LVBench) with 448 spatial resolution.
Specialized Metrics.
• OmniDocBench 1.5: Scores are computed as (1 − normalized Levenshtein distance) × 100, where higher values
indicate superior OCR and document understanding accuracy.
• WorldVQA: Access available at https://github.com/MoonshotAI/WorldVQA. This benchmark eval-
uates atomic, vision-centric world knowledge requiring fine-grained visual recognition and geographic under-
standing.
24
Kimi K2.5 T ECHNICAL R EPORT
E.5 Coding and Software Engineering
Terminal Bench 2.0. All scores are obtained using the default Terminus-2 agent framework with the provided JSON
parser. Notably, we evaluate under non-thinking mode because our current context management implementation for
thinking mode is technically incompatible with Terminus-2’s conversation state handling.
SWE-Bench Series. We employ an internally developed evaluation framework featuring a minimal tool set: bash,
create_file, insert, view, str_replace, and submit. System prompts are specifically tailored for
repository-level code manipulation. Peak performance is achieved under non-thinking mode across all SWE-Bench
variants (Verified, Multilingual, and Pro).
CyberGym. Claude Opus 4.5 results for this benchmark are reported under non-thinking settings as specified in
their technical documentation. We report scores in the difficulty level 1 (the primary setting).
PaperBench. We report the scores under the CodeDev setting.
Sampling. All coding task results are averaged over 5 independent runs (Avg@5) to ensure stability across environ-
ment initialization and non-deterministic test case ordering.
E.6 Agentic Evaluation
Tool Setting. Kimi-K2.5 is equipped with web search tool, code interpreter (Python execution environment), and
web browsing tools for all agentic evaluations, including HLE with tools and agentic search benchmarks (BrowseC-
omp, WideSearch, DeepSearchQA, FinSearchComp T2&T3 and Seal-0).
Context Management Strategies. To handle the extended trajectory lengths inherent in complex agentic tasks, we
implement domain-specific context management protocols. Unless otherwise specified below, no context manage-
ment is applied to agentic evaluations; tasks exceeding the model’s supported context window are directly counted as
failures rather than truncated.
• Humanity’s Last Exam (HLE). For the HLE tool-augmented setting, we employ a Hide-Tool-Result Context
Management strategy: when the context length exceeds predefined thresholds, only the most recent round of tool
messages (observations and return values) is retained, while the reasoning chain and thinking processes from all
previous steps are preserved in full.
• BrowseComp. For BrowseComp evaluations, our evaluation contains both with and without context management
settings. Under the context management setting, we adopt the same discard-all strategy proposed by DeepSeek,
where all history is truncated once token thresholds are exceeded.
System Prompt. All agentic search and HLE evaluations utilize the following unified system prompt, where DATE
is dynamically set to the current timestamp:
You are Kimi, today’s date: DATE.
Your task is to help the user with their questions by using various tools,
thinking deeply, and ultimately answering the user’s questions.
Please follow the following principles strictly during the deep research:
1. Always focus on the user’s original question during the research process,
avoiding deviating from the topic.
2. When facing uncertain information, use search tools to confirm.
3. When searching, filter high-trust sources (such as authoritative websites,
academic databases, and professional media) and maintain a critical mindset
towards low-trust sources.
4. When performing numerical calculations, prioritize using programming tools
to ensure accuracy.
5. Please use the format [^index^] to cite any information you use.
6. This is a **Very Difficult** problem--do not underestimate it. You must use
tools to help your reasoning and then solve the problem.
7. Before you finally give your answer, please recall what the question is
asking for.
25
Kimi K2.5 T ECHNICAL R EPORT
Sampling Protocol. To account for the inherent stochasticity in search engine result rankings and dynamic web
content availability, results for Seal-0 and WideSearch are averaged over 4 independent runs (Avg@4). All other
agentic benchmarks are evaluated under single-run protocols unless explicitly stated otherwise.
E.7 Computer-Use Evaluation
Hyperparameter Settings. We set max_steps_per_episode = 100 for all experiments, with
temperature = 0 for OSWorld-Verified and temperature = 0.1 for WebArena. Due to resource con-
straints, all models are evaluated in a one-shot setting. Adhering to the OpenCUA configuration [63], the agent
context includes the last 3 history images, the complete thought history, and the task instruction. For WebArena, we
manually corrected errors in the evaluation scripts and employed GPT-4o as the judge model for the fuzzy_match
function. To ensure fair comparison, Claude Opus 4.5 is evaluated solely with computer-use tools (excluding browser
tools), a departure from the System Card configuration [6].
System Prompt We utilize a unified system prompt for all computer use tasks:
You are a GUI agent. You are given an instruction, a screenshot of the screen and your
previous interactions with the computer. You need to perform a series of actions to
complete the task. The password of the computer is {password}.
For each step, provide your response in this format:
{thought}
## Action:
{action}
## Code:
{code}
In the code section, the code should be either pyautogui code or one of the following
functions wrapped in the code block:
- {"name": "computer.wait", "description": "Make the computer wait for 20 seconds
for installation, running code, etc.", "parameters": {"type": "object", "properties":
{}, "required": []}}
- {"name": "computer.terminate", "description": "Terminate the current task and report
its completion status", "parameters": {"type": "object", "properties": {"status":
{"type": "string", "enum": ["success", "failure"], "description": "The status of the
task"}, "answer": {"type": "string", "description": "The answer of the task"}},
"required": ["status"]}}
E.8 Agent Swarm Configuration
Tool Setting. In addition to the core toolset described in Appendix E.6 (web search, code interpreter, and web
browsing), the orchestrator is equipped with two specialized tools for sub-agent creation and scheduling:
• create_subagent: Instantiates a specialized sub-agent with a custom system prompt and identifier for reuse
across tasks.
• assign_task: Dispatches assignments to created sub-agents.
The tool schemas are provided below:
{
"name": "create_subagent",
"description": "Create a custom subagent with specific system prompt
and name for reuse.",
"parameters": {
"type": "object",
"properties": {
"name": {
"type": "string",
"description": "Unique name for this agent configuration"
},
"system_prompt": {
"type": "string",
26
Kimi K2.5 T ECHNICAL R EPORT
"description": "System prompt defining the agent’s role,
capabilities, and boundaries"
}
},
"required": ["name", "system_prompt"]
}
}
{
"name": "assign_task",
"description": "Launch a new agent.\nUsage notes:\n
1. You can launch multiple agents concurrently whenever possible,
to maximize performance;\n
2. When the agent is done, it will return a single message back to you.",
"parameters": {
"type": "object",
"properties": {
"agent": {
"type": "string",
"description": "Specify which created agent to use."
},
"prompt": {
"type": "string",
"description": "The task for the agent to perform"
}
},
"required": ["agent", "prompt"]
}
}
Step Limits. When operating in Agent Swarm mode, we set computational budgets for the orchestrator and sub-
agents. Step limits apply to the aggregate count of tool invocations and environment interactions.
• BrowseComp: The orchestrator is constrained to a maximum of 15 steps. Each spawned sub-agent operates
under a limit of 100 steps (i.e., up to 100 tool calls per sub-agent).
• WideSearch: Both the orchestrator and each sub-agent are allocated a maximum budget of 100 steps.
• In-house Bench: The orchestrator is constrained to a maximum of 100 steps. Each spawned sub-agent operates
under a limit of 50 steps .
System Prompt.
You are Kimi, a professional and meticulous expert in information collection and organization.
You fully understand user needs, skillfully use various tools, and complete tasks with the
highest efficiency.
# Task Description
After receiving users’ questions, you need to fully understand their needs and think
about and plan how to complete the tasks efficiently and quickly.
# Available Tools
To help you complete tasks better and faster, I have provided you with the following tools:
1. Search tool: You can use the search engine to retrieve information, supporting multiple
queries in parallel.
2. Browser tools: You can visit web links (web pages, PDFs, etc.), get page content, and
perform interactions such as clicking, inputting, finding, and scrolling.
3. Sub Agent tools:
- ‘create_subagent‘: Create a new sub-agent with a unique name and clear, specific
system prompt.
- ‘assign_task‘: Delegate tasks to created sub-agents. Sub-agents can also use search
and browser tools.
4. Other tools: Including code execution (IPython, Shell).
E.9 GDPVal
We cite the GDPVal-AA evaluation by Artificial Analysis, and the scores reported in Table 4 reflect the official leader-
board metrics as of January 28, 2026.
27
Kimi K2.5 T ECHNICAL R EPORT
USER
As a video web master, please analyze all these files (in ./video/wukong/) and create a cool HTML page:
1. Clarify the chronological timeline of video events
2. Identify the timestamps when player levels are upgraded.
32Parts, 24Hours, 40G
3. demonstrate your (k2.5 agent's) powerful long-video understanding and web-page design capability.
Output results to ./video/wukong-result-video-web-master/.
... ...
THINK SUB AGENT 32
<Think> The user wants me to analyze video files in `./video/wukong/` and create a cool HTML page that: ···
1. Clarifies the chronological timeline of video events
2. Identifies timestamps when player levels are upgraded SUB AGENT 3
3. Demonstrates my long-video understanding and web-page design capability
SUB AGENT 2
Let me first explore the directory to understand what files we're working with. </Think>
SUB AGENT 1
SUMMARY MAIN AGENT
I'll analyze the video files in the specified directory and create HTML page. Let me start by exploring what we have. Analyze this Black Myth: Wukong gameplay video: ./video/wukong/01-25895506462_Prologue.mp4
This is Video #01 - Prologue (320MB, ~5-8 minutes)
....
TOOL Save your analysis to:./video/wukong-result-video-web-master/analysis_01.json
ls -la ./video/wukong/ ....
THINK THINK
...
<Think> The user wants me to analyze a Black Myth: Wukong gameplay video
<Think> We have 32 video files totaling about 40GB! These are gameplay videos of "Black Myth: Wukong" - a
Chinese action RPG. Let me check the file sizes: ….
Let me start by exploring the file and extracting frames. </Think>
- Most files are 700MB - 2GB in size
- Total is 40GB+
SUMMARY
According to the AGENTS.md, for large videos (≥25MB), I should use subagents in parallel. Each subagent I'll analyze this Black Myth: Wukong gameplay video. Let me start by checking
··· the video and
should handle frame extraction and analysis independently. extracting key frames for analysis.
Let me create a strategy: TOOL
1. Each video is quite large (all >25MB), so I'll use subagents for each <Analyze Video> & <Extract Frames with FFMPEG> & <Read Media File x N>
2. Each subagent will analyze one video for:
- Timeline of events
- Level up timestamps ...
- Key moments/scenes
3. Then I'll aggregate the results and create an HTML page
Given the scale (32 large videos), I should spawn multiple subagents in parallel. Let me create subagents to SUMMARY
analyze each video. </Think>
## Analysis Complete ✅
SUMMARY I've successfully analyzed the Black Myth: Wukong Prologue video and created the structured
analysis file.
This is a massive project! 32 gameplay videos of "Black Myth: Wukong" (~40GB total). I'll deploy a parallel
……
analysis strategy using subagents for each video, then create a stunning HTML showcase.
**Files Created:**
- **Analysis JSON:** `./video/wukong-result-video-web-master/analysis_01.json`
TOOL - **Extracted Frames:** 82 frames at `./video/wukong-result-video-web-master/frames_01/`
<Create Subagent x 32>
THINK
<Think> Good, I can see all the analysis files are created. Now let me extract some key video clips/GIFs from important SUB AGENT 1
moments across the videos. I'll select key boss fights and level-up moments to create short clips for the HTML page. MAIN AGENT
Let me:
Extract a 10-second video clip from:
1. Extract some key video clips using ffmpeg
Source: ./video/wukong/01-25895506462_Prologue.mp4
2. Create the comprehensive HTML page
Time: 00:05:40 to 00:05:50
3. Let me first read a few analysis files to identify the best moments to extract. </Think>
Output: ./video/wukong-result-video-web-master/clips/01_prologue_battle.mp4
SUMMARY Use ffmpeg to extract this clip with high quality. This shows the epic tutorial battle with Erlang Shen
Let me read some analysis files to identify the best moments for video clips, then extract them in parallel starting.
Command format:
TOOL ffmpeg -ss 00:05:40 -t 10 -i ./video/wukong/01-25895506462_Prologue.mp4 -c:v libx264 -crf 23 -
<Read File x 32> preset fast -an ./video/wukong-result-video-web-master/clips/01_prologue_battle.mp4
Create the output directory first if needed.
THINK
<Think> Now I have a good understanding of the content. Let me extract some key video clips for the HTML
page. I'll extract:
1. Prologue - Epic battle with Erlang Shen (0:05:40-0:06:10)
THINK
...
<Think> The user wants me to extract a 10-second video clip from a specific video file using
ffmpeg. Let me first create the output directory if it doesn't exist, then run the ffmpeg
2. Chapter 1 boss fight - Black Wind King (0:27:00-0:27:30)
command to extract the clip. </Think>
……
6. Final Chapter - Flying on Somersault Cloud (0:12:30-0:13:00)
TOOL
Let me spawn subagents to extract these clips in parallel. </Think>
<Extract Video Clip with FFMPEG>
SUMMARY
Now let me extract key video clips for the HTML showcase. I'll spawn subagents to create short clips from the ...
most epic moments.
...
TOOL
THINK
<Create Subagent x 9>
The video clip was successfully extracted. Let me summarize the results for the main agent.
THINK
SUMMARY
<Think> I now have enough data. Let me also get a few more analysis files and then compile everything into a
✅ **Video clip extracted successfully!**
stunning HTML showcase. </Think>
**Output file:** `./video/wukong-result-video-web-master/clips/01_prologue_battle.mp4`
SUMMARY
Let me gather a few more key analysis files and then build the comprehensive HTML page. **Clip details:**
……
TOOL The 10-second clip showing the epic tutorial battle with Erlang Shen has been extracted with
<Build Frontend> high quality settings (CRF 23, fast preset) and saved to the output directory.
THINK
<Think> All tasks are complete. Let me provide a comprehensive summary to the user about what was accomplished. </Think>
SUMMARY
## ✅ Analysis Complete! Black Myth: Wukong Journey Showcase
……
### � How to View
Open `./video/wukong-result-video-web-master/index.html` in a web browser to experience the full interactive showcase with embedded video clips, animated charts, and the complete chronological timeline of the epic Black Myth:
Wukong journey!
Figure 11: Qualitative example of Kimi K2.5 analyzing a complete playthrough of Black Myth: Wukong (24 hours
of continuous gameplay across 32 videos at 1080p) using parallel visual agents. See generated webpage and source
videos (all rights reserved by source authors).
28
Kimi K2.5 T ECHNICAL R EPORT
USER USER USER
Please help the ant walk through the maze. Do not cross walls. What is the total percentage of the green and blue areas? Find the differences in th picture.
THINK THINK THINK
The user uploaded a maze image and asked me to help find a path from the start to the The user is asking for the total percentage of the green and blue areas in the image. This is a "Spot the Difference" game. The image shows the Flash the Sloth scene from
end without crossing walls, using the ipython tool for assistance. The image shows a pie chart with three parts: Section 1 (Red), Section 2 (Blue), and Zootopia.
First, I need to examine the maze image to determine the start and end positions: Section 3 (Green). Task: Analyze the differences between the top and bottom images.
Looking at the image, there is a small bell on the left, which is likely the start. Let me write code to: I need to load the image for pixel-level difference analysis.
There is an ant or insect character on the right, which is likely the end. 1. Load the image
The maze itself is an underground tunnel structure. 2. Identify the colored sectors
Let me use ipython to analyze this maze image and find the path. 3. Calculate the percentage of each color SUMMARY
Steps: 4. Sum green and blue Now conducting a pixel-wise comparison to find all differences accurately:
1. Analyze the maze structure (binarization, identify walls and paths). Actually, since this is a simple pie chart with distinct colors, I'll convert to HSV color
2. Identify the start and end points. space for better color segmentation, then mask each color and calculate the pixel TOOL
3. Use a pathfinding algorithm (like BFS or A*) to find the shortest path. counts.
<python: Difference map>
4. Visualize the path.^p
Let me start writing the code.
SUMMARY
I’ll help you analyze the percentage of the green and blue areas in this pie chart.
SUMMARY TOOL
The little bell on the left is the starting point, and the little ant on the right is the finish <python: Sample color algorithm>
line. Now, let me convert the maze into a computable format and find the path: # Convert to HSV for better color segmentation
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
# Define color ranges for each section… THINK
TOOL
# Red/Pink (Section 1) - need to handle red wrap-around in HSV Now I can see the obvious difference regions. Let me convert the contours into
<python: Binarization algorithm> lower_red1 = np.array([0, 50, 50]) bounding boxes and normalize the coordinates to the original image dimensions.
upper_red1 = np.array([10, 255, 255])
lower_red2 = np.array([160, 50, 50])
upper_red2 = np.array([180, 255, 255]) SUMMARY
Extracting bounding boxes from the detected differences and transforming them to
normalized coordinates.
TOOL
Total colored pixels: 61609
<python: Filter with threshold>
Red/Pink pixels: 26994
THINK
Blue pixels: 13752
The binarization process worked well. Now I need to: Green pixels: 18032
1. Manually mark the positions of the start (bell) and the end (ant). Sum of all colors: 58778
2. Implement the BFS algorithm to find the path.
Based on the image:
1. The starting point (bell) is on the left, roughly around (50, 250). THINK
2. The endpoint (ant) is on the right, roughly around (980, 450).
Good, I can see the masks are working but there's some discrepancy (58778 vs 61609
total pixels, probably due to black borders or anti-aliasing). Let me calculate the
SUMMARY percentages based on the sum of the three detected colors, which should give us the
relative proportions.
Now, let me determine the coordinates for the start and end points, and use the BFS
(Breadth-First Search) algorithm to find the shortest path: THINK
TOOL The automated detection identified 27 regions, but some may overlap or contain
TOOL multiple differences. I will manually refine the primary points of interest based on
<python: Visualization Results>
visual observation and provide more precise coordinates.
<python: Draw point and BFS> # Calculate percentages based on sum of detected colored pixels
total_detected = red_pixels + blue_pixels + green_pixels
red_pct = (red_pixels / total_detected) * 100 SUMMARY
blue_pct = (blue_pixels / total_detected) * 100 27 change regions were auto-detected, though many overlap. Let me manually refine
green_pct = (green_pixels / total_detected) * 100 these into the 8 major differences and output more accurate coordinates.
…
TOOL
<python: Filter main difference>
The path contains 3,288 points.
THINK
Great! The path has been found, consisting of 3,288 points. Now I need to visualize it
for the user. Let me create a colored path map and overlay the red path on the original
image.
SUMMARY
The entire route consists of 3,288 steps. Now, let me plot the path onto the maze:
TOOL THINK
Perfect! I've successfully analyzed the pie chart using image processing techniques. The
<python: Draw path>
results show:
Section 1 (Red/Pink): 45.9% THINK
Section 2 (Blue): 23.4% I have now identified all the differences and labeled their coordinates. Let me double-
Section 3 (Green): 30.7% check for any other obvious discrepancies.
Total of Green + Blue = 54.1% Upon further comparison, I also noticed:
Given the standard pie chart proportions, this looks like roughly: The sloth's hair/fur is different (smoother in the top image, more frizzy in the bottom).
Red: ~46% (close to 50% or half) There are subtle differences in the leaves of the background plants.
Blue: ~23% (close to 25% or a quarter) The paper next to the stamp on the desk.
Green: ~31% (close to 30%) However, the 10 main differences should cover the most prominent changes. Let me
So Green + Blue = ~54%, which is slightly more than half the pie chart. summarize the final answer.
The analysis is complete and I've provided both the detailed breakdown and the final
answer.
SUMMARY
THINK
SUMMARY After a thorough comparative analysis, a total of 10 major differences have been
Perfect! The path has been successfully found and visualized. identified.
Based on the image analysis using Python, here are the results:
Individual Percentages: 1. DMV vs EMV sign;
Section 1 (Red/Pink): 45.9% 2. Calendar image variation;
SUMMARY
Section 2 (Blue): 23.4% 3. Sloth expression: Closed vs Open mouth;
Final Path Characteristics: Section 3 (Green): 30.7% 4. Nick's eye state: Squinting vs Open;
🔍 Total Length: 3,288 pixels Answer to your question: The total percentage of the green and blue areas is: 5. Judy's facial adjustments;
🚶 Route: Enters from the left → heads down → traverses right across the underground 54.1% 6. Crumpled paper vs Paper airplane;
maze → winds upward → reaches the exit on the right. (Green 30.7% + Blue 23.4% = 54.1%) 7. Monitor display content;
🧠 Algorithm: BFS guarantees this is the shortest path with no backtracking. This means the green and blue sections combined make up slightly more than half of 8. Mug text change: 'YOU WANT IT WHEN?' vs 'I'LL GET TO IT';
🚧 Constraints: Strictly followed the corridors with no wall-clipping. TOOL
the entire pie chart. 9. Watch face/color variation;
The maze path is clearly marked in the image above; the red line represents the 10. File folder position/shape
<python: Color mask algorithm>
walking route!
Figure 12: Qualitative examples of Kimi K2.5 solving visual reasoning tasks via tool use.
29
Kimi K2.5 T ECHNICAL R EPORT
F Visualization
Figure 11 demonstrates our Agent Swarm tackling a challenging long-form video understanding task: analyzing a
complete playthrough of Black Myth: Wukong (24 hours of continuous gameplay across 32 videos, totaling 40GB).
The system employs a hierarchical multi-agent architecture where a Main Agent orchestrates parallel Sub Agents to
process individual video segments independently. Each sub agent performs frame extraction, temporal event analysis,
and key moment identification (e.g., boss fights, level-ups). The Main Agent subsequently aggregates these distributed
analyses to synthesize a comprehensive HTML showcase featuring chronological timelines, embedded video clips, and
interactive visualizations. This example demonstrates the system’s ability to handle massive-scale multimodal content
through parallelization while maintaining coherent long-context understanding.
Figure 12 presents qualitative examples of Kimi K2.5 solving diverse visual reasoning tasks via tool-augmented
reasoning. The model demonstrates: (1) Maze Solving—processing binary image segmentation and implementing
pathfinding algorithms (BFS) to navigate complex mazes; (2) Pie Chart Analysis—performing pixel-level color seg-
mentation and geometric calculations to determine precise area proportions; and (3) Spot-the-Difference—employing
computer vision techniques to detect pixel-level discrepancies between image pairs. These examples highlight the
model’s capability to decompose complex visual problems into executable code, iteratively refine strategies based on
intermediate results, and synthesize precise answers through quantitative visual analysis.
30
来源材料