DeepSeek-V3.2: Pushing the Frontier of Open
Large Language Models
DeepSeek-AI
research@deepseek.com
Abstract
arXiv:2512.02556v1 [cs.CL] 2 Dec 2025
We introduce DeepSeek-V3.2, a model that harmonizes high computational efficiency with supe-
rior reasoning and agent performance. The key technical breakthroughs of DeepSeek-V3.2 are as
follows: (1) DeepSeek Sparse Attention (DSA): We introduce DSA, an efficient attention mecha-
nism that substantially reduces computational complexity while preserving model performance
in long-context scenarios. (2) Scalable Reinforcement Learning Framework: By implementing
a robust reinforcement learning protocol and scaling post-training compute, DeepSeek-V3.2
performs comparably to GPT-5. Notably, our high-compute variant, DeepSeek-V3.2-Speciale,
surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro, achieving
gold-medal performance in both the 2025 International Mathematical Olympiad (IMO) and the
International Olympiad in Informatics (IOI). (3) Large-Scale Agentic Task Synthesis Pipeline:
To integrate reasoning into tool-use scenarios, we developed a novel synthesis pipeline that
systematically generates training data at scale. This methodology facilitates scalable agentic
post-training, yielding substantial improvements in generalization and instruction-following
robustness within complex, interactive environments.
DeepSeek-V3.2-Speciale DeepSeek-V3.2-Thinking GPT-5-High Claude-4.5-Sonnet Gemini-3.0-Pro
100 99.2 97.5 3000
96.0 94.6 95.0
93.1
90.2 2701 2708
87.0 88.3
2537 84.7 85.4
2386 80.380.2 2500
80 79.2
77.2 76.2
73.174.9
2000
Accuracy / Pass@1 (%) Codeforces Rating
60
54.2
1480
46.4
1500
42.8
40 37.7 38.6
36.4
35.2 35.2
30.6 1000
29.0
25.126.3
20
13.7
500
0 0
AIME 2025 HMMT 2025 HLE Codeforces SWE Terminal 2 Tool
(Pass@1) (Pass@1) (Pass@1) (Rating) Verified Bench 2.0 Bench Decathlon
(Resolved) (Acc) (Pass@1) (Pass@1)
Reasoning Capabilities Agentic Capabilities
Figure 1 | Benchmark of DeepSeek-V3.2 and its counterparts. For HMMT 2025, we report the
February competition, consistent with the baselines. For HLE, we report the text-only subset.
1. Introduction
The release of reasoning models (DeepSeek-AI, 2025; OpenAI, 2024a) marked a pivotal moment
in the evolution of Large Language Models (LLMs), catalyzing a substantial leap in overall
performance across the verifiable fields. Since this milestone, the capabilities of LLMs have
advanced rapidly. However, a distinct divergence has emerged in the past months. While
the open-source community (MiniMax, 2025; MoonShot, 2025; Qwen, 2025; ZhiPu-AI, 2025)
continues to make strides, the performance trajectory of closed-source proprietary models
(Anthropic, 2025b; DeepMind, 2025a; OpenAI, 2025) has accelerated at a significantly steeper rate.
Consequently, rather than converging, the performance gap between closed-source and open-
source models appears to be widening, with proprietary systems demonstrating increasingly
superior capabilities in complex tasks.
Through our analysis, we identify three critical deficiencies that limit the capability of open-
source models in complex tasks. First, architecturally, the predominant reliance on vanilla
attention (Vaswani et al., 2017) mechanisms severely constrains efficiency for long sequences.
This inefficiency poses a substantial obstacle to both scalable deployment and effective post-
training. Second, regarding resource allocation, open-source models suffer from insufficient
computational investment during the post-training phase, limiting their performance on hard
tasks. Finally, in the context of AI agents, open-source models demonstrate a marked lag in
generalization and instruction-following capabilities compared to their proprietary counterparts
(EvalSys, 2025; Li et al., 2025; Luo et al., 2025), hindering their effectiveness in real deployment.
To address these critical limitations, we first introduce DSA, a highly efficient attention
mechanism designed to substantially reduce computational complexity. This architecture
effectively addresses the efficiency bottleneck, preserving model performance even in long-
context scenarios. Second, we develop a stable and scalable RL protocol that allows for significant
computational expansion during the post-training phase. Notably, this framework allocates a
post-training computational budget exceeding 10% of the pre-training cost, unlocking advanced
capabilities. Thirdly, we propose a novel pipeline to foster generalizable reasoning in tool-use
scenarios. First, we implement a cold-start phase utilizing the DeepSeek-V3 (DeepSeek-AI,
2024) methodology to unify reasoning and tool-use within single trajectories. Subsequently, we
advance to large-scale agentic task synthesis, where we generate over 1,800 distinct environments
and 85,000 complex prompts. This extensive synthesized data drives the RL process, significantly
enhancing the model’s generalization and instruction-following capability in the agent context.
DeepSeek-V3.2 achieves similar performance with Kimi-k2-thinking and GPT-5 across mul-
tiple reasoning benchmarks. Furthermore, DeepSeek-V3.2 significantly advances the agentic
capabilities of open models, demonstrating exceptional proficiency on the long-tail agent tasks
introduced in EvalSys (2025); Li et al. (2025); Luo et al. (2025). DeepSeek-V3.2 emerges as a
highly cost-efficient alternative in agent scenarios, significantly narrowing the performance
gap between open and frontier proprietary models while incurring substantially lower costs.
Notably, with the aim of pushing the boundaries of open models in the reasoning domain, we
relaxed the length constraints to develop DeepSeek-V3.2-Speciale. As a result, DeepSeek-V3.2-
Speciale achieves performance parity with the leading closed-source system, Gemini-3.0-Pro
(DeepMind, 2025b). It shows gold-medal performance in the IOI 2025, ICPC World Final 2025,
IMO 2025, and CMO 2025.
2
2. DeepSeek-V3.2 Architecture
2.1. DeepSeek Sparse Attention
DeepSeek-V3.2 uses exactly the same architecture as DeepSeek-V3.2-Exp. Compared with
DeepSeek-V3.1-Terminus, the last version of DeepSeek-V3.1, the only architectural modification
of DeepSeek-V3.2 is the introduction of DeepSeek Sparse Attention (DSA) through continued
training.
Prototype of DSA. The prototype of DSA primarily consists of two components: a lightning
indexer and a fine-grained token selection mechanism.
The lightning indexer computes the index score 𝐼𝑡,𝑠 between the query token h𝑡 ∈ R𝑑 and a
preceding token h𝑠 ∈ R𝑑 , determining which tokens to be selected by the query token:
𝐻𝐼
∑︁
𝐼 𝑡 ,𝑠 = 𝑤𝑡𝐼, 𝑗 · ReLU q𝑡𝐼, 𝑗 · k𝑠𝐼 , (1)
𝑗=1
where 𝐻 𝐼 denotes the number of indexer heads; q𝑡𝐼, 𝑗 ∈ R𝑑 and 𝑤𝑡𝐼, 𝑗 ∈ R are derived from the
𝐼
query token h𝑡 ; and k𝑠𝐼 ∈ R𝑑 is derived from the preceding token h𝑠 . We choose ReLU as the
𝐼
activation function for throughput consideration. Given that the lightning indexer has a small
number of heads and can be implemented in FP8, its computational efficiency is remarkable.
Given the index scores { 𝐼𝑡,𝑠 } for each query token h𝑡 , our fine-grained token selection
mechanism retrieves only the key-value entries {c𝑠 } corresponding to the top-k index scores.
Then, the attention output u𝑡 is computed by applying the attention mechanism between the
query token h𝑡 and the sparsely selected key-value entries {c𝑠 }:
u𝑡 = Attn h𝑡 , c𝑠 𝐼𝑡,𝑠 ∈ Top-k 𝐼𝑡,: . (2)
Instantiate DSA Under MLA. For the consideration of continued training from DeepSeek-
V3.1-Terminus, we instantiate DSA based on MLA (DeepSeek-AI, 2024) for DeepSeek-V3.2. At
the kernel level, each key-value entry must be shared across multiple queries for computational
efficiency (Yuan et al., 2025). Therefore, we implement DSA based on the MQA (Shazeer, 2019)
mode of MLA1 , where each latent vector (the key-value entry of MLA) will be shared across
all query heads of the query token. The DSA architecture based on MLA is illustrated in
Figure 2. We also provide an open-source implementation of DeepSeek-V3.22 to specify the
details unambiguously.
2.1.1. Continued Pre-Training
Starting from a base checkpoint of DeepSeek-V3.1-Terminus, whose context length has been ex-
tended to 128K, we perform continued pre-training followed by post-training to create DeepSeek-
V3.2.
The continued pre-training of DeepSeek-V3.2 consists of two training stages. For both stages,
the distribution of training data is totally aligned with the 128K long context extension data
used for DeepSeek-V3.1-Terminus.
1We illustrate the difference between the MQA and MHA modes of MLA in Appendix A.
2 https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp/tree/main/inference
3
Output Hidden 𝐮𝐮𝑡𝑡 ··· ···
{𝐨𝐨𝐶𝐶𝑡𝑡,𝑖𝑖 } {𝐨𝐨𝑡𝑡,𝑖𝑖 }
···
Multi-Query Attention (Core Attention)
𝐴𝐴
{[𝐪𝐪𝑡𝑡,𝑖𝑖 ; 𝐪𝐪𝑅𝑅𝑡𝑡,𝑖𝑖 ]} Top-k Selector
···
concatenate
Lightning
𝐴𝐴
{𝐪𝐪𝑡𝑡,𝑖𝑖 } {[𝐜𝐜𝑡𝑡𝐾𝐾𝐾𝐾 ; 𝐤𝐤 𝑅𝑅𝑡𝑡 ]} ···
··· Indexer
concatenate
{𝐪𝐪𝑅𝑅𝑡𝑡,𝑖𝑖 }
{𝐪𝐪𝐶𝐶𝑡𝑡,𝑖𝑖 } apply RoPE {𝐪𝐪𝐼𝐼𝑡𝑡,𝑗𝑗 } 𝐤𝐤 𝐼𝐼𝑡𝑡 𝐼𝐼
{𝑤𝑤𝑡𝑡,𝑗𝑗 }
𝐤𝐤 𝑅𝑅𝑡𝑡 𝐜𝐜𝑡𝑡𝐾𝐾𝐾𝐾 ···
partially partially
apply RoPE
𝑄𝑄 apply RoPE apply RoPE
𝐜𝐜𝑡𝑡 ···
Input Hidden 𝐡𝐡𝑡𝑡 ··· ···
Figure 2 | Attention architecture of DeepSeek-V3.2, where DSA is instantiated under MLA. The
green part illustrates how DSA selects the top-k key-value entries according to the indexer.
Dense Warm-up Stage. We first use a short warm-up stage to initialize the lightning indexer.
In this stage, we keep dense attention and freeze all model parameters except for the lightning
indexer. To align the indexer outputs with the main attention distribution, for the 𝑡 -th query
token, we first aggregate the main attention scores by summing across all attention heads.
This sum is then L1-normalized along the sequence dimension to produce a target distribution
𝑝𝑡,: ∈ R𝑡 . Based on 𝑝𝑡,: , we set a KL-divergence loss as the training objective of the indexer:
∑︁
DKL 𝑝𝑡,: Softmax 𝐼𝑡,: .
L𝐼 = (3)
𝑡
For warm-up, we use a learning rate of 10−3 . We train the indexer for only 1000 steps, with each
step consisting of 16 sequences of 128K tokens, resulting in a total of 2.1B tokens.
Sparse Training Stage. Following indexer warm-up, we introduce the fine-grained token
selection mechanism and optimize all model parameters to adapt the model to the sparse
pattern of DSA. In this stage, we also keep aligning the indexer
outputs to the main attention
distribution, but considering only the selected token set S𝑡 = 𝑠 𝐼𝑡,𝑠 ∈ Top-k 𝐼𝑡,: :
∑︁
DKL 𝑝𝑡,S𝑡 Softmax 𝐼𝑡,S𝑡 .
L𝐼 = (4)
𝑡
It is worth noting that we detach the indexer input from the computational graph for separate
optimization. The training signal of the indexer is from only L 𝐼 , while the optimization of the
main model is according to only the language modeling loss. In this sparse training stage, we
use a learning rate of 7.3 × 10−6 , and select 2048 key-value tokens for each query token. We train
both the main model and the indexer for 15000 steps, with each step consisting of 480 sequences
of 128K tokens, resulting in a total of 943.7B tokens.
4
2.2. Parity Evaluation
Standard Benchmark In September 2025, we evaluate DeepSeek-V3.2-Exp on a suite of bench-
marks, which focus on diverse capabilities, and compare it with DeepSeek-V3.1-Terminus
showing similar performance. While DeepSeek V3.2 Exp significantly improves computational
efficiency on long sequences, we do not observe substantial performance degradation compared
with DeepSeek-V3.1-Terminus, on both short- and long-context tasks.
Human Preference Given that direct human preference assessments are inherently suscep-
tible to bias, we employ ChatbotArena as an indirect evaluation framework to approximate
user preferences for the newly developed base models. Both DeepSeek-V3.1-Terminus and
DeepSeek-V3.2-Exp share an identical post-training strategy, and their Elo scores, obtained from
evaluations conducted on 10 November 2025, are closely matched. These results suggest that the
new base model achieves performance on par with the previous iteration, despite incorporating
a sparse attention mechanism.
Long Context Eval Following the release of DeepSeek-V3.2-Exp, several independent long-context
evaluations were conducted using previously unseen test sets. A representative benchmark
is AA-LCR3 , in which DeepSeek-V3.2-Exp scores four points higher than DeepSeek-V3.1-
Terminus in reasoning mode. In the Fiction.liveBench evaluation4 , DeepSeek-V3.2-Exp consis-
tently outperforms DeepSeek-V3.1-Terminus across multiple metrics. This evidence indicates
the base checkpoint of DeepSeek-V3.2-Exp does not regress on long context tasks.
2.3. Inference Costs
DSA reduces the core attention complexity of the main model from O 𝐿2 to O ( 𝐿𝑘), where 𝑘
(≪ 𝐿) is the number of selected tokens. Although the lightning indexer still has a complexity
of O 𝐿2 , it requires much less computation compared with MLA in DeepSeek-V3.1-Terminus.
Combined with our optimized implementation, DSA achieves a significant end-to-end speedup
in long-context scenarios. Figure 3 presents how token costs of DeepSeek-V3.1-Terminus and
DeepSeek-V3.2 vary with the token position in the sequence. These costs are estimated from
benchmarking the actual service deployed on H800 GPUs, at a rental price of 2 USD per GPU
hour. Note that for short-sequence prefilling, we specially implement a masked MHA mode to
simulate DSA, which can achieve higher efficiency under short-context conditions.
3. Post-Training
After continued pre-training, we perform post-training to create the final DeepSeek-V3.2. The
post-training of DeepSeek-V3.2 also employs sparse attention in the same way as the sparse
continued pre-training stage. For DeepSeek-V3.2, we maintain the same post-training pipeline
as in DeepSeek-V3.2-Exp, which includes specialist distillation and mixed RL training.
Specialist Distillation For each task, we initially develop a specialized model dedicated
exclusively to that particular domain, with all specialist models being fine-tuned from the same
3 https://artificialanalysis.ai/evaluations/artificial-analysis-long-context-reasoni
ng
4 https://fiction.live/stories/Fiction-liveBench-April-6-2025/oQdzQvKHw8JyXbN87
5
0.7$ 2.4$
DeepSeek-V3.1-Terminus DeepSeek-V3.1-Terminus
0.6$ DeepSeek-V3.2 2.0$ DeepSeek-V3.2
Cost Per Million Tokens Cost Per Million Tokens
0.5$
1.6$
0.4$
1.2$
0.3$
0.8$
0.2$
0.1$ 0.4$
0$ 0$
0K 32K 64K 96K 128K 0K 32K 64K 96K 128K
Token Position Token Position
(a) Prefilling (b) Decoding
Figure 3 | Inference costs of DeepSeek-V3.1-Terminus and DeepSeek-V3.2 on H800 clusters.
pre-trained DeepSeek-V3.2 base checkpoint. In addition to writing tasks and general question-
answering, our framework encompasses six specialized domains: mathematics, programming,
general logical reasoning, general agentic tasks, agentic coding, and agentic search, with all the
domains supporting both thinking and non-thinking modes. Each specialist is trained with large-
scale Reinforcement Learning (RL) computing. Furthermore, we employ different models to
generate training data for long chain-of-thought reasoning (thinking mode) and direct response
generation (non-thinking mode). Once the specialist models are prepared, they are used to
produce the domain-specific data for the final checkpoint. Experimental results demonstrate that
models trained on the distilled data achieve performance levels only marginally below those
of domain-specific specialists, with the performance gap being effectively eliminated through
subsequent RL training.
Mixed RL Training For DeepSeek-V3.2, we still adopt Group Relative Policy Optimization
(GRPO) (DeepSeek-AI, 2025; Shao et al., 2024) as the RL training algorithm. As DeepSeek-
V3.2-Exp, we merge reasoning, agent, and human alignment training into one RL stage. This
approach effectively balances performance across diverse domains while circumventing the
catastrophic forgetting issues commonly associated with multi-stage training paradigms. For
reasoning and agent tasks, we employ rule-based outcome reward, length penalty, and language
consistency reward. For general tasks, we employ a generative reward model where each
prompt has its own rubrics for evaluation.
DeepSeek-V3.2 and DeepSeek-V3.2-Speciale DeepSeek-V3.2 integrates reasoning, agent, and
human alignment data distilled from specialists, undergoing thousands of steps of continued RL
training to reach the final checkpoints. To investigate the potential of extended thinking, we also
developed an experimental variant, DeepSeek-V3.2-Speciale. This model was trained exclusively
on reasoning data with a reduced length penalty during RL. Additionally, we incorporated the
dataset and reward method from DeepSeekMath-V2 (Shao et al., 2025) to enhance capabilities in
mathematical proofs.
We would like to highlight our efforts in how to create a stable recipe to scale up RL compute
in Section 3.1, and how to integrate thinking into agentic tasks in Section 3.2
6
3.1. Scaling GRPO
We first review the objective of GRPO. GRPO optimizes the policy model 𝜋𝜃 by maximizing the
following objective on a group of responses { 𝑜1 , · · · , 𝑜𝐺 } sampled from the old policy 𝜋old given
each question 𝑞:
|𝑜 |
" 𝐺 𝑖
1 ∑︁ 1 ∑︁
JGRPO ( 𝜃) = E𝑞∼𝑃 ( 𝑄 ),{ 𝑜𝑖 } 𝐺 ∼𝜋old (· | 𝑞 )
𝑖=1 𝐺 | 𝑜𝑖 |
𝑖=1 𝑡 =1
#
min 𝑟𝑖,𝑡 ( 𝜃) 𝐴ˆ 𝑖,𝑡 , clip 𝑟𝑖,𝑡 ( 𝜃), 1 − 𝜀, 1 + 𝜀 𝐴ˆ 𝑖,𝑡 − 𝛽 DKL 𝜋𝜃 ( 𝑜𝑖,𝑡 ) 𝜋ref ( 𝑜𝑖,𝑡 ) ,
(5)
where
𝜋𝜃 ( 𝑜𝑖,𝑡 | 𝑞, 𝑜𝑖,<𝑡 )
𝑟𝑖,𝑡 ( 𝜃) = (6)
𝜋old ( 𝑜𝑖,𝑡 | 𝑞, 𝑜𝑖,<𝑡 )
is the importance sampling ratio between the current and old policy. 𝜀 and 𝛽 are hyper-
parameters controlling the clipping range and KL penalty strength, respectively. 𝐴ˆ 𝑖,𝑡 is the
advantage of 𝑜𝑖,𝑡 which is estimated by normalizing the outcome reward within each group.
Specifically, a set of reward models are used to score an outcome reward 𝑅 𝑖 for each output
𝑜𝑖 in the group, yielding 𝐺 rewards 𝑹 = { 𝑅1 , · · · , 𝑅𝐺 } respectively. The advantage of 𝑜𝑖,𝑡 is
calculated by subtracting the average reward of the group from the reward of output 𝑜𝑖 , i.e.,
𝐴ˆ 𝑖,𝑡 = 𝑅 𝑖 − mean( 𝑹).
In the following, we outline additional strategies that stabilize RL scaling, directly building
on the GRPO algorithm.
Unbiased KL Estimate Given 𝑜𝑖,𝑡 is sampled from the old policy 𝜋old (·| 𝑞, 𝑜𝑖,<𝑡 ), we correct the
K3 estimator (Schulman, 2020) to obtain an unbiased KL estimate using the importance-sampling
ratio between the current policy 𝜋𝜃 and the old policy 𝜋old .
𝜋𝜃 ( 𝑜𝑖,𝑡 | 𝑞, 𝑜𝑖,<𝑡 ) 𝜋ref ( 𝑜𝑖,𝑡 | 𝑞, 𝑜𝑖,<𝑡 ) 𝜋ref ( 𝑜𝑖,𝑡 | 𝑞, 𝑜𝑖,<𝑡 )
DKL 𝜋𝜃 ( 𝑜𝑖,𝑡 ) 𝜋ref ( 𝑜𝑖,𝑡 ) =
− log − 1 . (7)
𝜋old ( 𝑜𝑖,𝑡 | 𝑞, 𝑜𝑖,<𝑡 ) 𝜋𝜃 ( 𝑜𝑖,𝑡 | 𝑞, 𝑜𝑖,<𝑡 ) 𝜋𝜃 ( 𝑜𝑖,𝑡 | 𝑞, 𝑜𝑖,<𝑡 )
As a direct result of this adjustment, the gradient of this KL estimator becomes unbiased,
which eliminates systematic estimation errors, thereby facilitating stable convergence. This
contrasts sharply with the original K3 estimator, particularly when the sampled tokens have
substantially lower probabilities under the current policy than the reference policy, i.e., 𝜋𝜃 ≪ 𝜋ref .
In such cases, the gradient of the K3 estimator assigns disproportionately large, unbounded
weights to maximize the likelihood of these tokens, resulting in noisy gradient updates that
accumulate to degrade sample quality in subsequent iterations and lead to unstable training
dynamics. In practice, we find that different domains benefit from varying strengths of KL
regularization. For certain domains, such as mathematics, applying a relatively weak KL penalty
or even omitting it entirely can yield improved performance.
Off-Policy Sequence Masking To improve the efficiency of RL systems, we typically generate
a large batch of rollout data, which is subsequently split into multiple mini-batches for several
gradient update steps. This practice inherently introduces off-policy behavior. Additionally,
inference frameworks used for efficient data generation are often highly optimized, which may
differ in implementation details from training frameworks. Such training-inference inconsistency
7
further exacerbates the degree of off-policyness. To stabilize training and improve tolerance for
off-policy updates, we mask negative sequences that introduce significant policy divergence, as
measured by the KL divergence between the data-sampling policy 𝜋old and the current policy
𝜋𝜃 . More specifically, we introduce a binary mask 𝑀 into the GRPO loss:
|𝑜 |
" 𝐺 𝑖
1 ∑︁ 1 ∑︁
JGRPO ( 𝜃) = E𝑞∼𝑃 ( 𝑄 ),{ 𝑜𝑖 } 𝐺 ∼𝜋old (· | 𝑞 )
𝑖=1 𝐺 | 𝑜𝑖 |
𝑖=1 𝑡 =1
#
min 𝑟𝑖,𝑡 ( 𝜃) 𝐴ˆ 𝑖,𝑡 , clip 𝑟𝑖,𝑡 ( 𝜃), 1 − 𝜀, 1 + 𝜀 𝐴ˆ 𝑖,𝑡 𝑀𝑖,𝑡 − 𝛽 DKL 𝜋𝜃 ( 𝑜𝑖,𝑡 ) 𝜋ref ( 𝑜𝑖,𝑡 ) ,
(8)
where
Í|𝑜 | 𝜋 ( 𝑜𝑖,𝑡 | 𝑞,𝑜𝑖,<𝑡 )
0 𝐴ˆ 𝑖,𝑡 < 0, | 𝑜1𝑖 | 𝑡=1𝑖 log 𝜋old
>𝛿
𝜃 ( 𝑜𝑖,𝑡 | 𝑞,𝑜𝑖,<𝑡 )
𝑀 𝑖 ,𝑡 = (9)
1 otherwise,
and 𝛿 is a hyper-parameter that controls the threshold of policy divergence. Note that 𝜋old
here denotes the sampling probability directly returned by the inference framework, thus the
KL divergence between the old and current policy accounts for both sources of off-policyness
mentioned above. It is also worth noting that we only mask sequences with negative advantages.
Intuitively, the model benefits the most by learning from its own mistakes, whereas highly
off-policy negative samples can be detrimental, potentially misleading or destabilizing the
optimization process. We empirically observe that this Off-Policy Sequence Masking operation
improves stability in certain training scenarios that would otherwise exhibit instability.
Keep Routing Mixture-of-Experts (MoE) models improve computational efficiency by activat-
ing only a subset of expert modules during inference. However, discrepancies between inference
and training frameworks, compounded by policy updates, can result in inconsistent expert
routing during inference and training even for identical inputs. Such inconsistency induces
abrupt shifts in the active parameter subspace, which destabilizes optimization and exacerbates
off-policy issues. To mitigate this, we preserve the expert routing paths used during sampling
in the inference framework and enforce the same routing paths during training, ensuring that
identical expert parameters are optimized. This Keep Routing operation was found crucial for
RL training stability of MoE models, and has been adopted in our RL training pipeline since
DeepSeek-V3-0324.
Keep Sampling Mask Top-p and top-k sampling are widely used sampling strategies to
enhance the quality of responses generated by LLMs. Employing these strategies in RL training
is also advantageous, as it avoids sampling extremely low-probability tokens that would be
used as optimization targets. While such truncation preserves sample quality, it introduces a
mismatch between the action spaces of 𝜋old and 𝜋𝜃 , which violates the principles of importance
sampling and destabilizes training. To address this, we preserve the truncation masks during
sampling from 𝜋old and apply them to 𝜋𝜃 during training, ensuring both policies share identical
action subspaces. Empirically, we find that combining top-p sampling with the Keep Sampling
Mask strategy effectively preserves language consistency during RL training.
8
3.2. Thinking in Tool-Use
3.2.1. Thinking Context Management
DeepSeek-R1 has demonstrated that incorporating a thinking process can significantly enhance
a model’s ability to solve complex problems. Building on this insight, we aim to integrate
thinking capabilities into tool-calling scenarios.
We observed that replicating DeepSeek-R1’s strategy—discarding reasoning content upon the
arrival of the second round of messages—results in significant token inefficiency. This approach
forces the model to redundantly re-reason through the entire problem for each subsequent
tool call. To mitigate this, we developed a context management strictly tailored for tool-calling
scenarios as shown in Fig 4:
• Historical reasoning content is discarded only when a new user message is introduced
to the conversation. If only tool-related messages (e.g., tool outputs) are appended, the
reasoning content is retained throughout the interaction.
• When reasoning traces are removed, the history of tool calls and their results remains
preserved in the context.
Notably, certain agent frameworks, such as Roo Code or Terminus, simulate tool interactions
via user messages. These frameworks may not fully benefit from our enhanced reasoning
persistence due to the context management rules outlined above. Therefore, we recommend
utilizing non-thinking models for optimal performance with such architectures.
Figure 4 | Thinking retention mechanism in tool-calling scenarios.
3.2.2. Cold-Start
Given the availability of reasoning data (non-agentic) and non-reasoning agentic data, a straight-
forward strategy for integrating these two capabilities is through carefully designed prompting.
We posit that the model possesses sufficient ability to accurately follow explicit instructions,
thereby enabling the seamless incorporation of tool execution within the reasoning process.
9
To demonstrate the operation of the cold-start mechanism, we selectively sample the training
data as shown in Appendix Tables 6–8. It is important to note that distinct task prompts are
associated with different system prompts. Tables 6–8 present an illustrative example correspond-
ing to a competitive programming prompt. Table 6 presents an example of our reasoning data,
which uses a system prompt to explicitly asks the model to do reasoning before the final answer
and uses a special tag <think></think> to label the reasoning path. Table 7 shows the prompt
of non-reasoning agentic data, where the system prompt contains the guidance of toolcall. Table
8 presents the system prompt we designed to instruct the model to incorporate multiple tool
calls within its reasoning process.
In this manner, although the reasoning in tool-use patterns may lack robustness, the model
is occasionally able to generate the desired trajectories, thereby providing a basis for subsequent
reinforcement learning stages.
3.2.3. Large-Scale Agentic Tasks
A diverse set of RL tasks is crucial for enhancing model robustness. For tasks such as search,
code engineering, and code interpretation, we employ real-world tools, including actual web
search APIs, coding tools, and Jupyter Notebooks. While these RL environments are real, the
prompts employed are either extracted from Internet sources or synthetically generated, rather
than obtained from actual user interactions. For other tasks, the environment and prompts are
both synthetically constructed. The agent tasks we used are described in Table 1.
Table 1 | The description of different agent tasks, including the number of tasks, environment
type (real or synthesized), and prompt source (extracted or synthesized).
number of tasks environment prompt
code agent 24667 real extracted
search agent 50275 real synthesized
general agent 4417 synthesized synthesized
code interpreter 5908 real extracted
Search Agent We employ a multi-agent pipeline based on DeepSeek-V3.2 to generate diverse,
high-quality training data. We first sample informative long-tail entities across diverse domains
from large-scale web corpora. A question-construction agent then explores each entity using
search tools with configurable depth and breadth parameters, consolidating the discovered
information into question-answer pairs. Multiple answer-generation agents with heteroge-
neous configurations (different checkpoints, system prompts, etc.) produce diverse candidate
responses for each proposed QA pair. A verification agent with search capabilities validates
all answers through multiple passes, retaining only samples where the ground-truth is correct
and all candidates are verifiably incorrect. These data spans multiple languages, domains, and
difficulty levels. To complement these verifiable samples and better reflect real-world usage,
we also augment the dataset with filtered instances from our existing helpful RL datasets, for
which the search tool provides measurable benefits. We then develop detailed evaluation rubrics
across multiple quality dimensions and employ a generative reward model to score responses
based on these rubrics. This hybrid approach enables optimization for both factual reliability
and practical helpfulness.
10
Code Agent We constructed large-scale, executable environments for software issue resolution
by mining millions of issue-Pull Request (PR) pairs from GitHub. This dataset was rigorously
filtered using heuristic rules and LLM-based judgments to ensure high quality, requiring that
each entry contain a reasonable issue description, a correlated gold patch, and a test patch for
validation. An automated environment-setup agent, powered by DeepSeek-V3.2, was employed
to build executable environments for these pairs. This agent handles package installation, de-
pendency resolution, and test execution. Test results are output in the standard JUnit format,
ensuring consistent parsing across programming languages and test frameworks. An environ-
ment is deemed successfully built only when applying the gold patch results in a non-zero count
of false-to-positive (F2P) test cases (indicating the issue is fixed) and a zero count of pass-to-fail
(P2F) test cases (indicating no regressions). Using this pipeline, we successfully built tens of
thousands of reproducible issue resolution environments spanning multiple programming
languages, including Python, Java, JavaScript, TypeScript, C, C++, Go, and PHP.
Code Interpreter Agent We utilize Jupyter Notebook as a code interpreter to address complex
reasoning tasks. To facilitate this, we curate a diverse set of problems spanning mathematics,
logic, and data science, each requiring the model to leverage code execution capabilities to arrive
at a solution.
General Agent To scale up agent environments and tasks in RL, we employ an automatic
environment-synthesis agent that synthesizes 1,827 task-oriented environments. These tasks are
hard to solve but easy to verify. The synthesis workflow primarily consists of environment and
toolset construction, task synthesis, and solution generation. Specifically, the workflow proceeds
as follows.
1. Given a task category (e.g., planning a travel itinerary) and a sandbox equipped with a
bash and a search tool, the agent first uses these tools to generate or retrieve relevant data
from the Internet and store them in the sandbox database.
2. The agent then synthesizes a set of task-specific tools, each implemented as a function.
3. To create tasks that are both challenging and automatically verifiable, the agent initially
proposes a simple task based on the current database, along with its solution and verifica-
tion functions implemented in Python. The solution function is restricted to invoking tool
functions or performing logical computations, and cannot call other functions or directly
access the database, ensuring the task can only be solved through the tool interface. Addi-
tionally, the results produced by the solution function must be validated by the verification
function. If the solution is not validated, the agent will modify the solution or verification
functions until the solution’s output passes the verification. The agent then iteratively
increases the difficulty of the task and updates the corresponding solution and verification
functions. During this iterative process, if the current toolset is not sufficient to solve the
task, the agent will augment the toolset.
Following this workflow, we obtain thousands of <environment, tools, task, verifier> tuples.
We then perform RL on this dataset using DeepSeek-V3.2 and retain only instances with non-zero
pass@100, resulting in 1,827 environments and their corresponding tasks (4,417 in total). A syn-
thetic trip-planning example is illustrated below. This example highlights that, while searching
the large combinatorial space for a trip plan that satisfies all constraints is challenging, checking
whether a given candidate solution satisfies these constraints is relatively straightforward.
11
An Example of Synthesized Task: Trip Planning
I’m planning a three-day trip starting from Hangzhou, and I need help creating an itinerary
from October 1st to October 3rd, 2025. A few important requirements: I don’t want to repeat
any cities, hotels, attractions, or restaurants during the entire trip. Also, please make sure that
every hotel, restaurant, and attraction you recommend is actually located in the city where
I’ll be staying that day. One more thing about the second day - I’m trying to be smart about
my budget. If I end up booking a luxury hotel that costs 800 CNY or more per night, then I
need to be more careful with other expenses: my total spending on both restaurants (lunch
and dinner) should stay under 350 CNY, both restaurants should be rated at least 4.0 stars,
and the afternoon attraction ticket needs to be less than 120 CNY. If the hotel on day 2 is in
the mid-to-high range (500-800 CNY), then I have a bit more flexibility - I just need to make
sure at least one of my restaurant choices is rated 4.0 or higher, and the attraction ticket should
be below 180 CNY. For more affordable hotels (200-500 CNY range), I only need to ensure
that at least one restaurant has a rating of 3.2 or above. Can you help me put together this itinerary?
Submit Result Format
[
{ "time": "2025-10-01", "city": "cite_name", "hotel": "hotel_name", "afternoon_restaurant": "restau-
rant_name", "afternoon_attraction": "attraction_name", "evening_restaurant": "restaurant_name" },
{ "time": "2025-10-02", "city": "cite_name", "hotel": "hotel_name", "afternoon_restaurant": "restau-
rant_name", "afternoon_attraction": "attraction_name", "evening_restaurant": "restaurant_name" },
{ "time": "2025-10-03", "city": "cite_name", "hotel": "hotel_name", "afternoon_restaurant": "restau-
rant_name", "afternoon_attraction": "attraction_name", "evening_restaurant": "restaurant_name" }
]
Tool Set for Trip Planning
Function Name Description
get_all_attractions_by_city(city) Get all attractions for given city.
get_all_cities() Get all cities from the database.
get_all_hotels_by_city(city) Get all hotels for given city.
get_all_restaurants_by_city(city) Get all restaurants for given city.
get_city_by_attraction(attraction) Get city for given attraction name.
get_city_by_hotel(hotel) Get city for given hotel name.
get_city_by_restaurant(restaurant) Get city for given restaurant name.
get_city_transport(city) Get all intra-city transport options for given city.
get_infos_by_attraction(info_keywords, attraction) Get specified infos for given attraction.
get_infos_by_city(info_keywords, city) Get specified infos for given city.
get_infos_by_hotel(info_keywords, hotel) Get specified infos for given hotel.
get_infos_by_restaurant(info_keywords, restaurant) Get specified infos for given restaurant.
get_inter_city_transport(from_city, to_city) Get all transports between given city pair.
get_weather_by_city_date(city, date) Get weather for given city-date pair.
submit_result(answer_text) Submit the final answer content.
4. Evaluation
4.1. Main Results
We evaluate models on MMLU-Pro (Wang et al., 2024), GPQA Diamond (Rein et al., 2023),
Human Last Exam (HLE) Text-only (Phan et al., 2025), LiveCodeBench (2024.08-2025.04), Code-
12
forces, Aider-Polyglot, AIME 2025, HMMT Feb 2025, HMMT Nov 2025 (Balunović et al., 2025),
IMOAnswerBench (Luong et al., 2025), Terminal Bench 2.0, SWE-Verified (OpenAI, 2024b), SWE
Multilingual (Yang et al., 2025), BrowseComp (Wei et al., 2025), BrowseCompZh (Zhou et al.,
2025), 𝜏2 -bench (Barres et al., 2025), MCP-Universe (Luo et al., 2025), MCP-Mark (EvalSys, 2025),
and Tool-Decathlon (Li et al., 2025). Tool-use benchmarks are evaluated using the standard
function call format, wherein models are configured to thinking mode. For MCP-Universe
(Luo et al., 2025) and MCP-Mark (EvalSys, 2025), we evaluate all models with our internal
environment, because the search and playwright environment might be slightly different from
the official setting. We set the temperature to 1.0, and the context window to 128K tokens.
For math-related tasks such as AIME, HMMT, IMOAnswerBench, and HLE, we eval with
the following template: "{question}\nPlease reason step by step, and put your
final answer within \boxed{}." In the case of HLE, we additionally assessed DeepSeek-
V3.2-Thinking using the official template, resulting in a score of 23.9.
Table 2 | Comparison between DeepSeek-V3.2 and closed/open models. For open models,
we just compare with models supports thinking in tooluse. Numbers in bold represent the
best scores within each model class (open-source and closed-source). The 𝜏2 -Bench result is
computed by the average of each category. Regarding BrowseComp, the performance with the
context management technique is noted with *.
Claude-4.5- GPT-5 Gemini-3.0 Kimi-K2 MiniMax DeepSeek-V3.2
Benchmark (Metric)
Sonnet High Pro Thinking M2 Thinking
MMLU-Pro (EM) 88.2 87.5 90.1 84.6 82.0 85.0
English GPQA Diamond (Pass@1) 83.4 85.7 91.9 84.5 77.7 82.4
HLE (Pass@1) 13.7 26.3 37.7 23.9 12.5 25.1
LiveCodeBench (Pass@1-COT) 64.0 84.5 90.7 82.6 83.0 83.3
Code
Codeforces (Rating) 1480 2537 2708 - - 2386
AIME 2025 (Pass@1) 87.0 94.6 95.0 94.5 78.3 93.1
HMMT Feb 2025 (Pass@1) 79.2 88.3 97.5 89.4 - 92.5
Math
HMMT Nov 2025 (Pass@1) 81.7 89.2 93.3 89.2 - 90.2
IMOAnswerBench (Pass@1) - 76.0 83.3 78.6 - 78.3
Terminal Bench 2.0 (Acc) 42.8 35.2 54.2 35.7 30.0 46.4
Code Agent SWE Verified (Resolved) 77.2 74.9 76.2 71.3 69.4 73.1
SWE Multilingual (Resolved) 68.0 55.3 - 61.1 56.5 70.2
BrowseComp (Pass@1) 24.1 54.9 - -/60.2* 44.0 51.4/67.6*
Search Agent BrowseCompZh (Pass@1) 42.4 63.0 - 62.3 48.5 65.0
HLE (Pass@1) 32.0 35.2 45.8 44.9 31.8 40.8
𝜏2 -Bench(Pass@1) 84.7 80.2 85.4 74.3 76.9 80.3
MCP-Universe (Success Rate) 46.5 47.9 50.7 35.6 29.4 45.9
ToolUse
MCP-Mark (Pass@1) 33.3 50.9 43.1 20.4 24.4 38.0
Tool-Decathlon (Pass@1) 38.6 29.0 36.4 17.6 16.0 35.2
DeepSeek-V3.2 achieves similar performance with GPT-5-high on reasoning tasks, but is
slightly worse than Gemini-3.0-Pro. Compared to K2-Thinking, DeepSeek-V3.2 achieves compa-
rable scores with substantially fewer output tokens, as shown in Table 3. These performance
gains can be attributed to the increased computational resources allocated to RL training. Over
recent months, we have observed consistent performance improvements correlating with ex-
tended RL training budget, which already exceeds 10% of the pre-training cost. We hypothesize
that reasoning capabilities could be further enhanced with additional computational budget al-
location. Notably, the performance of DeepSeek-V3.2 presented herein is constrained by a length
constraint reward model; upon removal of the restriction, we observe further improvement in
13
model performance, as detailed in Section 4.2.
In code agent evaluations, DeepSeek-V3.2 significantly outperforms open-source LLMs on
both SWE-bench Verified and Terminal Bench 2.0, demonstrating its potential within real-world
coding workflows. Regarding Terminal Bench 2.0, as previously noted, our context management
strategy for the ’thinking mode’ is currently incompatible with Terminus; consequently, the
reported score of 46.4 was achieved using the Claude Code framework. We also evaluated
DeepSeek-V3.2 with Terminus in non-thinking mode, yielding a score of 39.3. For SWE-bench
Verified, the primary score was obtained using our internal framework. Robustness tests across
other settings—including the Claude Code and RooCode frameworks, as well as non-thinking
mode—produced consistent results, ranging from 72 to 74.
For the search agent evaluation, we assess our models using a standard commercial search
API. Since DeepSeek-V3.2 supports a maximum context length of only 128K, approximately
20%+ of the test cases exceed this limit. To address this, we employ a context management
method to derive the final score. For reference, the score is 51.4 without context management.
Further details are provided in Section 4.4.
On tool-use benchmarks, DeepSeek-V3.2 substantially narrows the performance gap between
open-source and closed-source LLMs, though it remains below frontier models. For 𝜏2 -bench,
we employ the model itself as the user agent, achieving final category scores of 63.8 (Airline),
81.1 (Retail), and 96.2 (Telecom). For the MCP benchmarks, we employ the function calling
format and place tool outputs within messages designated with the ’tool’ role, rather than
the ’user’ role. During our testing, we observed that DeepSeek-V3.2 frequently engages in
redundant self-verification, generating excessively long trajectories. This tendency often causes
the context length to exceed the 128K limit, particularly in tasks such as MCP-Mark GitHub
and Playwright evaluation. Consequently, this phenomenon hinders the final performance
of DeepSeek-V3.2. However, integrating context management strategies can further enhance
performance. We identify this as a direction for future work and a practical consideration for
users. Even if DeepSeek-V3.2 suffers from the issue, it still significantly outperforms existing
open models. Notably, since the environments and toolsets employed in these benchmarks were
not encountered during RL training, the observed improvements demonstrate DeepSeek-V3.2’s
capacity to generalize its reasoning strategies to out-of-domain agentic scenarios. The evaluation
of non-thinking model in the agent scenario is shown in Appendix Table 9.
4.2. Results of DeepSeek-V3.2-Speciale
Table 3 demonstrates that DeepSeek-V3.2-Speciale achieves superior performance by leveraging
increased reasoning tokens, surpassing the state-of-the-art Gemini-3.0-Pro across multiple
benchmarks. Remarkably, as shown in Table 4, this general-purpose model attains gold-medal
level performance in the 2025 International Olympiad in Informatics (IOI) and the ICPC World
Finals (ICPC WF) without targeted training. Furthermore, by incorporating techniques from
Shao et al. (2025), the model excels in complex proof tasks, reaching gold-medal thresholds
in the 2025 International Mathematical Olympiad (IMO) and China Mathematical Olympiad
(CMO)5 . Detailed evaluation protocols are provided in Appendix D.
However, the token efficiency of DeepSeek-V3.2-Speciale remains significantly inferior to
that of Gemini-3.0-Pro. To mitigate deployment costs and latency, we imposed stricter token
constraints during the training of the official DeepSeek-V3.2, aiming to optimize the trade-off
5We evaluated the English version of CMO 2025. The IMO 2025 and CMO 2025 problems, together with the
inference code, can be found at: https://github.com/deepseek-ai/DeepSeek-Math-V2.
14
Table 3 | Benchmark performance and efficiency of reasoning models. For each benchmark, cells
show accuracy and output token count (in thousands). The highest accuracy per benchmark is
in bold; the second-highest is underlined.
GPT-5 Gemini-3.0 Kimi-K2 DeepSeek-V3.2 DeepSeek-V3.2
Benchmark
High Pro Thinking Thinking Speciale
AIME 2025 (Pass@1) 94.6 (13k) 95.0 (15k) 94.5 (24k) 93.1 (16k) 96.0 (23k)
HMMT Feb 2025 (Pass@1) 88.3 (16k) 97.5 (16k) 89.4 (31k) 92.5 (19k) 99.2 (27k)
HMMT Nov 2025 (Pass@1) 89.2 (20k) 93.3 (15k) 89.2 (29k) 90.2 (18k) 94.4 (25k)
IMOAnswerBench (Pass@1) 76.0 (31k) 83.3 (18k) 78.6 (37k) 78.3 (27k) 84.5 (45k)
LiveCodeBench (Pass@1-COT) 84.5 (13k) 90.7 (13k) 82.6 (29k) 83.3 (16k) 88.7 (27k)
CodeForces (Rating) 2537 (29k) 2708 (22k) - 2386 (42k) 2701 (77k)
GPQA Diamond (Pass@1) 85.7 (8k) 91.9 (8k) 84.5 (12k) 82.4 (7k) 85.7 (16k)
HLE (Pass@1) 26.3 (15k) 37.7 (15k) 23.9 (24k) 25.1 (21k) 30.6 (35k)
between performance and cost. We believe that token efficiency remains a critical area for future
investigation.
Table 4 | Performance of DeepSeek-V3.2-Speciale in top-tier mathematics and coding compe-
titions. For ICPC WF 2025, we report the number of submissions for each successfully solved
problem. DeepSeek-V3.2-Speciale ranked 2nd in ICPC WF 2025 and 10th in IOI 2025.
Competition P1 P2 P3 P4 P5 P6 Overall Medal
IMO 2025 7 7 7 7 7 0 35/42 Gold
CMO 2025 18 18 9 21 18 18 102/126 Gold
IOI 2025 100 82 72 100 55 83 492/600 Gold
Competition A B C D E F G H I J K L Overall Medal
ICPC WF 2025 3 - 1 1 2 2 - 1 1 1 1 1 10/12 Gold
4.3. Synthesis Agentic Tasks
In this section, we perform ablation experiments to study the effect of synthetic agentic tasks.
We focus on two questions. First, are synthetic tasks sufficiently challenging for reinforcement
learning? Second, how well do these synthetic tasks generalize, i.e., can they transfer to different
downstream tasks or real-world environments?
To address the first question, we randomly sample 50 instances from the general synthesized
agentic tasks and evaluate both the model used for synthesis and frontier closed-source LLMs.
As shown in Table 5, DeepSeek-V3.2-Exp attains an accuracy of only 12%, while frontier closed-
source models achieve at most 62%. These results indicate that the synthetic data include agentic
tasks that are challenging for both DeepSeek-V3.2-Exp and frontier closed-source models.
To investigate whether RL on synthetic data can generalize to different tasks or real-world
environments, we apply RL to the SFT checkpoint of DeepSeek-V3.2 (denoted DeepSeek-V3.2-
SFT). To exclude the effects of long CoT and other RL data, we conduct RL only on synthetic
agentic tasks in non-thinking mode. We then compare the model with DeepSeek-V3.2-SFT
and DeepSeek-V3.2-Exp, where DeepSeek-V3.2-Exp is trained with RL only in search and code
environments. As shown in Figure 5, large-scale RL on synthetic data yields substantial improve-
15
Table 5 | Accuracy of general synthesized tasks on different models.
Pass@K DeepSeek-v3.2-Exp Sonnet-4.5 Gemini-3.0 Pro GPT-5-Thinking
1 12% 34% 51% 62%
2 18% 47% 65% 75%
4 26% 62% 74% 82%
Figure 5 | RL training of DeepSeek-V3.2-SFT using exclusively synthetic general agent data.
ments over DeepSeek-V3.2-SFT on Tau2Bench, MCP-Mark, and MCP-Universe benchmarks. In
contrast, restricting RL to code and search scenarios does not improve performance on these
benchmarks, further highlighting the potential of synthetic data.
4.4. Context Management of Search Agent
Even with extended context windows such as 128k, agentic workflows, particularly in search-
based scenarios, frequently encounter maximum length limitations that prematurely truncate
the reasoning process. This bottleneck inhibits the full realization of test-time compute potential.
To address this, we introduce context management employing simple strategies to extend token
budgets at test time，when the token usage exceeds 80% of the context window length. These
strategies include (1) Summary, which summarizes the overflowed trajectory and re-initiates
the rollout; (2) Discard-75%, which discards the first 75% tool call history in the trajectory to
free up spaces; (3) Discard-all, which resets the context by discarding all previous tool call
history (similar to the new context tool (Anthropic, 2025a)). For comparison, we also implement
a parallel scaling baseline, Parallel-fewest-step, which samples N independent trajectories and
16
70.0
Summary
Discard-75%
67.5 Discard-all
Parallel-fewest-step
65.0
62.5
Browsecomp
60.0
57.5
55.0
52.5
100 200 300 400 500 600 700 800 900
Real Steps
Figure 6 | Accuracy of Browsecomp with different test-time compute expansion strategies.
selects the trajectory with the fewest steps.
We evaluate these strategies on the BrowseComp benchmark (Wei et al., 2025). As illustrated
in Figure 6, under varying compute budgets, context management leads to significant perfor-
mance gains by allowing the model to scale up test-time compute, providing more space to
perform additional execution steps. For example, Summary extends the average steps to 364,
achieving a performance improvement of up to 60.2. However, its overall efficiency is relatively
low. Despite its simplicity, Discard-all performs well in both efficiency and scalability, achieving
a score of 67.6, comparable to parallel scaling while using significantly fewer steps.
In summary, test-time compute can be scaled either serially through context management
or in parallel, both effectively extending the model’s problem-solving capacity. However,
different strategies exhibit varying efficiency and scalability. Thus, it is crucial to account for
actual compute costs when benchmarking model performance. Meanwhile, finding the optimal
combination of serial and parallel scaling to maximize both efficiency and scalability remains a
crucial direction for future work.
5. Conclusion, Limitation, and Future Work
In this work, we introduced DeepSeek-V3.2, a framework that effectively bridges the gap be-
tween computational efficiency and advanced reasoning capabilities. Using DSA, we addressed
critical computation complexity without sacrificing long-context performance. By increasing
computational budget, DeepSeek-V3.2 achieves comparable performance with GPT-5 on rea-
soning benchmarks. Finally, the integration of our large-scale agentic task synthesis pipeline
significantly enhances tool-use proficiency, unlocking new possibilities for robust and generaliz-
able AI agents with open LLM. Furthermore, our high-compute variant, DeepSeek-V3.2-Speciale,
validated by gold-medal achievements in the IMO and IOI, sets a milestone for open LLMs.
Despite these achievements, we acknowledge certain limitations when compared to frontier
closed-source models such as Gemini-3.0-Pro. First, due to fewer total training FLOPs, the
breadth of world knowledge in DeepSeek-V3.2 still lags behind that of leading proprietary
17
models. We plan to address this knowledge gap in future iterations by scaling up the pre-training
compute. Second, token efficiency remains a challenge; DeepSeek-V3.2 typically requires longer
generation trajectories (i.e., more tokens) to match the output quality of models like Gemini-
3.0-Pro. Future work will focus on optimizing the intelligence density of the model’s reasoning
chains to improve efficiency. Third, solving complex tasks is still inferior to frontier models,
motivating us to further refine our foundation model and post-training recipe.
References
Anthropic. System card: Claude opus 4.5, 2025a. URL https://assets.anthropic.com/m
/64823ba7485345a7/Claude-Opus-4-5-System-Card.pdf.
Anthropic. Introducing claude sonnet 4.5, 2025b. URL https://www.anthropic.com/news
/claude-sonnet-4-5l.
M. Balunović, J. Dekoninck, I. Petrov, N. Jovanović, and M. Vechev. Matharena: Evaluating llms
on uncontaminated math competitions. Proceedings of the Neural Information Processing
Systems Track on Datasets and Benchmark, 2025.
V. Barres, H. Dong, S. Ray, X. Si, and K. Narasimhan. 𝜏2 -bench: Evaluating conversational agents
in a dual-control environment, 2025. URL https://arxiv.org/abs/2506.07982.
DeepMind. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long
context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025a.
G. DeepMind. Gemini 3 pro model card, 2025b. URL https://storage.googleapis.com
/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf.
DeepSeek-AI. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language
model. CoRR, abs/2405.04434, 2024. doi: 10.48550/ARXIV.2405.04434. URL https:
//doi.org/10.48550/arXiv.2405.04434.
DeepSeek-AI. Deepseek-v3 technical report, 2024. URL https://arxiv.org/abs/2412.194
37.
DeepSeek-AI. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning.
Nature, 645(8081):633–638, 2025.
EvalSys. Mcpmark leaderboard, 2025. URL https://mcpmark.ai/leaderboard.
J. Li, W. Zhao, J. Zhao, W. Zeng, H. Wu, X. Wang, R. Ge, Y. Cao, Y. Huang, W. Liu, et al. The
tool decathlon: Benchmarking language agents for diverse, realistic, and long-horizon task
execution. arXiv preprint arXiv:2510.25726, 2025.
Z. Luo, Z. Shen, W. Yang, Z. Zhao, P. Jwalapuram, A. Saha, D. Sahoo, S. Savarese, C. Xiong, and
J. Li. Mcp-universe: Benchmarking large language models with real-world model context
protocol servers. arXiv preprint arXiv:2508.14704, 2025.
T. Luong, D. Hwang, H. H. Nguyen, G. Ghiasi, Y. Chervonyi, I. Seo, J. Kim, G. Bingham,
J. Lee, S. Mishra, A. Zhai, C. H. Hu, H. Michalewski, J. Kim, J. Ahn, J. Bae, X. Song, T. H.
Trinh, Q. V. Le, and J. Jung. Towards robust mathematical reasoning. In Proceedings of
the 2025 Conference on Empirical Methods in Natural Language Processing, 2025. URL
https://aclanthology.org/2025.emnlp-main.1794/.
18
MiniMax. https://www.minimax.io/news/minimax-m2, 2025. URL https://www.minimax.
io/news/minimax-m2.
MoonShot. Introducing kimi k2 thinking, 2025. URL https://moonshotai.github.io/Kim
i-K2/thinking.html.
OpenAI. Learning to reason with llms, 2024a. URL https://openai.com/index/learnin
g-to-reason-with-llms/.
OpenAI. Introducing SWE-bench verified we’re releasing a human-validated subset of swe-
bench that more, 2024b. URL https://openai.com/index/introducing-swe-bench
-verified/.
OpenAI. Introducing gpt-5, 2025. URL https://openai.com/index/introducing-gpt-5
/.
L. Phan, A. Gatti, Z. Han, N. Li, J. Hu, H. Zhang, C. B. C. Zhang, M. Shaaban, J. Ling, S. Shi, et al.
Humanity’s last exam. arXiv preprint arXiv:2501.14249, 2025.
Qwen. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388.
D. Rein, B. L. Hou, A. C. Stickland, J. Petty, R. Y. Pang, J. Dirani, J. Michael, and S. R. Bowman.
GPQA: A graduate-level google-proof q&a benchmark. arXiv preprint arXiv:2311.12022, 2023.
J. Schulman. Approximating KL divergence, 2020. URL http://joschu.net/blog/kl-app
rox.html.
Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, M. Zhang, Y. K. Li, Y. Wu, and D. Guo. Deepseek-
math: Pushing the limits of mathematical reasoning in open language models. CoRR,
abs/2402.03300, 2024. doi: 10.48550/ARXIV.2402.03300. URL https://doi.org/10
.48550/arXiv.2402.03300.
Z. Shao, Y. Luo, C. Lu, Z. Ren, J. Hu, T. Ye, Z. Gou, S. Ma, and X. Zhang. Deepseekmath-v2:
Towards self-verifiable mathematical reasoning, 2025.
N. Shazeer. Fast transformer decoding: One write-head is all you need. CoRR, abs/1911.02150,
2019. URL http://arxiv.org/abs/1911.02150.
A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polo-
sukhin. Attention is all you need. pages 5998–6008, 2017. URL https://proceedings.neur
ips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html.
Y. Wang, X. Ma, G. Zhang, Y. Ni, A. Chandra, S. Guo, W. Ren, A. Arulraj, X. He, Z. Jiang, T. Li,
M. Ku, K. Wang, A. Zhuang, R. Fan, X. Yue, and W. Chen. Mmlu-pro: A more robust and
challenging multi-task language understanding benchmark. CoRR, abs/2406.01574, 2024.
URL https://doi.org/10.48550/arXiv.2406.01574.
J. Wei, Z. Sun, S. Papay, S. McKinney, J. Han, I. Fulford, H. W. Chung, A. T. Passos, W. Fedus,
and A. Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents. arXiv
preprint arXiv:2504.12516, 2025.
J. Yang, K. Lieret, C. E. Jimenez, A. Wettig, K. Khandpur, Y. Zhang, B. Hui, O. Press, L. Schmidt,
and D. Yang. Swe-smith: Scaling data for software engineering agents, 2025. URL https:
//arxiv.org/abs/2504.21798.
19
J. Yuan, H. Gao, D. Dai, J. Luo, L. Zhao, Z. Zhang, Z. Xie, Y. Wei, L. Wang, Z. Xiao, Y. Wang,
C. Ruan, M. Zhang, W. Liang, and W. Zeng. Native sparse attention: Hardware-aligned
and natively trainable sparse attention. In W. Che, J. Nabende, E. Shutova, and M. T. Pile-
hvar, editors, Proceedings of the 63rd Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers), ACL 2025, pages 23078–23097. Association for Compu-
tational Linguistics, 2025. URL https://aclanthology.org/2025.acl-long.1126/.
ZhiPu-AI. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models. arXiv preprint
arXiv:2508.06471, 2025.
P. Zhou, B. Leon, X. Ying, C. Zhang, Y. Shao, Q. Ye, D. Chong, Z. Jin, C. Xie, M. Cao, et al.
Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese.
arXiv preprint arXiv:2504.19314, 2025.
Appendices
A. MHA and MQA Modes of MLA
Output Hidden 𝐮! ··· ···
Output Hidden 𝐮! ··· ···
𝑾𝑼𝑽 '
𝒊 𝐨$,&
{𝐨$!,# } {𝐨!,# }
{𝐨!,# } ···
Multi-Head Attention (Core Attention)
Multi-Query Attention (Core Attention)
)
{[𝐪!,# ; 𝐪(!,# ]}
{[𝐪$!,# ; 𝐪(!,# ]} {[𝐤 $!,# ; 𝐤 (! ]} ···
concatenate {[𝐜!%& ; 𝐤 (! ]} ···
concatenate concatenate
) concatenate
{𝐪!,# }
$
···
{𝐪$!,# } {𝐪(!,# } 𝐤 (! {𝐤 $!,# } {𝐯!,# } {𝐪(!,# }
𝑾𝑼𝑲 '
𝒊 𝐪$,&
apply RoPE
𝑾𝑼𝑲 %&
𝒊 𝐜$ 𝑾𝑼𝑽 %&
𝒊 𝐜$ {𝐪$!,# } apply RoPE 𝐤 (! 𝐜!%& ···
apply RoPE
'
··· 𝐜! 𝐜!%& ··· ' apply RoPE
··· 𝐜!
Input Hidden 𝐡! ··· ··· Input Hidden 𝐡! ··· ···
(a) MHA mode of MLA. (b) MQA mode of MLA.
Figure 7 | Illustration of the MHA and MQA modes of MLA. For DeepSeek-V3.1-Terminus, the
MHA mode is used for training and prefilling, while the MQA mode is used for decoding.
Figure 7 illustrates two aspects of MLA – the MHA and MQA modes – as well as the
transformation between them.
B. Cold Start Template
20
Table 6 | An example of the reasoning data system prompt. The system prompt requires the
model to output the reasoning process in the tag <think></think>.
Reasoning You are an expert Python programmer. You will be given a question (problem
System specification) and will generate a correct Python program that matches the spec-
Prompt ification and passes all tests. Please first reason before giving the final answer.
The reasoning process enclosed within <think> </think>. The final answer is
output after the </think> tag.
Prompt Given a linked list, swap every two adjacent nodes and return its head ...
Reasoning <think>
Response ...
</think>
[FINAL ANSWER]
Table 7 | {TOOL-DESCRIPTIONS} and {TOOLCALL-FORMAT} will be replaced with the specific
tools and our designed toolcall format.
Agent Use Python interpreter tool to execute Python code. The code will not be shown
System to the user. This tool should be used for internal reasoning, but not for code that
Prompt is intended to be visible to the user (e.g. when creating plots, tables, or files).
When you send a message containing Python code to python, it will be executed
in a stateful Jupyter notebook environment. python will respond with the output
of the execution or time out after 120.0 seconds.
## Tools
You have access to the following tools:
{TOOL-DESCRIPTIONS}
Important: ALWAYS adhere to this exact format for tool use:
{TOOLCALL-FORMAT}
Prompt Given a linked list, swap every two adjacent nodes and return its head ...
Agent Re- [MULTI-TURN TOOLCALL]
sponse [FINAL ANSWER]
Table 8 | The model executes tool calls in thinking process.
Reasoning You are a helpful assistant with access to a Python interpreter.
Required - You may use the Python tool **multiple times** during your reasoning, a.k.a in
Agent <think></think>, with a maximum of 20 code executions.
System - Call the Python tool early in your reasoning to aid in solving the task. Continue
Prompt reasoning and invoking tools as needed until you reach the final answer. Once
you have the answer, stop reasoning and present your solution using Markdown
and LaTeX.
- Do NOT invoke any tools in your presented final solution steps.
- To improve efficiency and accuracy, you should prefer code execution over
language-based reasoning whenever possible. Keep your reasoning succinct; let
the code do the heavy lifting.
## Tools
You have access to the following tools:
{TOOL-DESCRIPTIONS}
Important: ALWAYS adhere to this exact format for tool use:
{TOOLCALL-FORMAT}
Prompt Given a linked list, swap every two adjacent nodes and return its head ...
Agent <think>
Response [MULTI-TURN Thinking-Then-TOOLCALL]
with </think>
Thinking [FINAL ANSWER]
21
C. Non-thinking DeepSeek-V3.2 Agentic Evaluation
Table 9 | Comparison between DeepSeek-V3.2 non-thinking and thinking modes. The terminal
bench scores are evaluated with the Claude Code framework in the table. Non-thinking score of
Terminal Bench 2.0 with Terminus framework is 39.3.
Benchmark (Metric) non-thinking thinking
Terminal Bench 2.0 (Acc) 37.1 46.4
Code Agent SWE Verified (Resolved) 72.1 73.1
SWE Multilingual (Resolved) 68.9 70.2
𝜏2 -bench (Pass@1) 77.2 80.3
MCP-Universe (Success Rate) 38.6 45.9
ToolUse
MCP-Mark (Pass@1) 26.5 38.0
Tool-Decathlon (Pass@1) 25.6 35.2
The performance of non-thinking mode is slightly worse than the thinking mode, but still
competitive.
D. Evaluation Method of IOI, ICPC World Final, IMO, and CMO
For all competitions, the model’s maximum generation length is set to 128k. No tools or internet
access are used, and testing strictly adheres to the contest’s time and attempt limits.
For the IOI evaluation, we designed our submission strategy in accordance with the official
competition rules, which permit up to 50 submissions per problem and score each submission
based on the maximum points achieved across all subtasks. Specifically, we first sampled 500
candidate solutions for each problem, then applied a multi-stage filtering pipeline. In the initial
stage, we eliminated invalid submissions that failed to pass the provided sample test cases or
exceeded the length constraints. Subsequently, we employed the DeepSeek-V32-Exp model to
identify and remove samples in which the model explicitly indicated an inability or refusal to
solve the problem. From the remaining valid candidates, we selected the 50 samples with the
longest thinking traces for final submission.
For the ICPC evaluation, we adapted the same filtering methodology but with a smaller
initial sample size. We generated 32 candidate solutions per problem and applied the identical
filtering criteria to select submissions.
In the IMO and CMO tasks, we employ a generate-verify-refine loop. The model iteratively
improves its solution until it achieves a perfect self-evaluation or hits the maximum revision
cap, identical to the process in Shao et al. (2025).
22
E. Author List
Research & Engineering: Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang,
Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chengda Lu, Chenggang
Zhao, Chengqi Deng, Chenhao Xu, Chong Ruan*, Damai Dai, Daya Guo, Dejian Yang, Deli
Chen, Erhang Li, Fangqi Zhou*, Fangyun Lin, Fucong Dai, Guangbo Hao, Guanting Chen,
Guowei Li, H. Zhang, Hanwei Xu, Hao Li, Haofen Liang, Haoran Wei, Haowei Zhang, Haowen
Luo, Haozhe Ji, Honghui Ding, Hongxuan Tang, Huanqi Cao, Huazuo Gao, Hui Qu, Hui Zeng,
Jialiang Huang, Jiashi Li, Jiaxin Xu, Jiewen Hu, Jingchang Chen, Jingting Xiang, Jingyang Yuan,
Jingyuan Cheng, Jinhua Zhu, Jun Ran*, Junguang Jiang, Junjie Qiu, Junlong Li*, Junxiao Song,
Kai Dong, Kaige Gao, Kang Guan, Kexin Huang*, Kexing Zhou, Kezhao Huang, Kuai Yu,
Lean Wang, Lecong Zhang, Lei Wang, Liang Zhao, Liangsheng Yin*, Lihua Guo, Lingxiao Luo,
Linwang Ma, Litong Wang, Liyue Zhang, M.S. Di, M.Y Xu, Mingchuan Zhang, Minghua Zhang,
Minghui Tang, Mingxu Zhou, Panpan Huang, Peixin Cong, Peiyi Wang, Qiancheng Wang,
Qihao Zhu, Qingyang Li, Qinyu Chen, Qiushi Du, Ruiling Xu, Ruiqi Ge, Ruisong Zhang, Ruizhe
Pan, Runji Wang, Runqiu Yin, Runxin Xu, Ruomeng Shen, Ruoyu Zhang, S.H. Liu, Shanghao Lu,
Shangyan Zhou, Shanhuang Chen, Shaofei Cai, Shaoyuan Chen, Shengding Hu, Shengyu Liu,
Shiqiang Hu, Shirong Ma, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, Songyang
Zhou, Tao Ni, Tao Yun, Tian Pei, Tian Ye, Tianyuan Yue, Wangding Zeng, Wen Liu, Wenfeng
Liang, Wenjie Pang, Wenjing Luo, Wenjun Gao, Wentao Zhang, Xi Gao, Xiangwen Wang, Xiao
Bi, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaokang Zhang, Xiaotao Nie, Xin Cheng,
Xin Liu, Xin Xie, Xingchao Liu, Xingkai Yu, Xingyou Li, Xinyu Yang, Xinyuan Li*, Xu Chen,
Xuecheng Su, Xuehai Pan, Xuheng Lin, Xuwei Fu, Y.Q. Wang, Yang Zhang, Yanhong Xu, Yanru
Ma, Yao Li, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Qian, Yi Yu, Yichao Zhang, Yifan
Ding, Yifan Shi, Yiliang Xiong, Ying He, Ying Zhou, Yinmin Zhong, Yishi Piao, Yisong Wang,
Yixiao Chen, Yixuan Tan, Yixuan Wei, Yiyang Ma, Yiyuan Liu, Yonglun Yang, Yongqiang Guo,
Yongtong Wu, Yu Wu, Yuan Cheng, Yuan Ou, Yuanfan Xu, Yuduan Wang, Yue Gong*, Yuhan
Wu, Yuheng Zou, Yukun Li, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang
Zhou, Z.F. Wu, Z.Z. Ren, Zehua Zhao, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda
Xie, Zhengyan Zhang, Zhewen Hao, Zhibin Gou, Zhicheng Ma, Zhigang Yan, Zhihong Shao,
Zhixian Huang, Zhiyu Wu, Zhuoshu Li, Zhuping Zhang, Zian Xu, Zihao Wang, Zihui Gu, Zijia
Zhu, Zilin Li, Zipeng Zhang, Ziwei Xie, Ziyi Gao, Zizheng Pan, Zongqing Yao
Data Annotation: Bei Feng, Hui Li, J.L. Cai, Jiaqi Ni, Lei Xu, Meng Li, Ning Tian, R.J. Chen,
R.L. Jin, S.S. Li, Shuang Zhou, Tianyu Sun, X.Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen,
Xinnan Song, Xinyi Zhou, Y.X. Zhu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian
Ma, Zhen Huang, Zhipeng Xu, Zhongyu Zhang
Business & Compliance: Dongjie Ji, Jian Liang, Jianzhong Guo, Jin Chen, Leyi Xia, Miaojun
Wang, Mingming Li, Peng Zhang, Ruyi Chen, Shangmian Sun, Shaoqing Wu, Shengfeng Ye,
T.Wang, W.L. Xiao, Wei An, Xianzu Wang, Xiaowen Sun, Xiaoxiang Wang, Ying Tang, Yukun
Zha, Zekai Zhang, Zhe Ju, Zhen Zhang, Zihua Qu
Authors are listed alphabetically by their first name. Names marked with * denote individu-
als who have departed from our team.
23
Source notes and reports