Paper 2512.02556
DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
- Published
- Dec 2025
- Research lab
- DeepSeek
- Citations
- 671
- GitHub
- Not linked
01 In brief
Summary
DeepSeek-V3.2 introduces three key innovations to close the gap with closed-source models: DeepSeek Sparse Attention (DSA) reduces computational complexity from O(L^2) to O(Lk) while preserving long-context performance; a scalable reinforcement learning (RL) framework with a post-training compute budget exceeding 10% of pre-training cost enables performance comparable to GPT-5; and a large-scale agentic task synthesis pipeline generates over 1,800 environments and 85,000 prompts to enhance tool-use generalization.
The high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and matches Gemini-3.0-Pro, achieving gold-medal performance at IMO 2025, IOI 2025, ICPC WF 2025, and CMO 2025.
DSA uses a lightning indexer and fine-grained token selection, trained via a dense warm-up (2.1B tokens) and sparse training (943.7B tokens).
RL stability is improved with unbiased KL estimates, off-policy sequence masking, and keep routing/sampling masks.
Context management retains reasoning during tool calls and discards it only on new user messages.
On benchmarks, DeepSeek-V3.2 scores 93.1 on AIME 2025, 2386 Codeforces rating, 73.1 on SWE Verified, and 80.3 on Tau2-Bench, while Speciale achieves 96.0 AIME, 2701 Codeforces, and 99.2 HMMT Feb 2025.
Limitations include lower world knowledge breadth and token efficiency compared to frontier models.
02 From the paper
Abstract
We introduce DeepSeek-V3.2, a model that harmonizes high computational efficiency with superior 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 mechanism 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.