Elamur Research
☰
home / papers / 2025 / DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv — non-exclusive licence to distribute" does not allow redistributing the text itself. Original — arxiv.org/pdf/2512.02556

Summary

Open-weight LLMs have been falling further behind proprietary systems. The authors trace this to three causes: dense attention is costly on long inputs, post-training receives little compute, and agentic generalization is weak. DeepSeek-V3.2 tackles each one. First, it adds DeepSeek Sparse Attention (DSA) to DeepSeek-V3.1-Terminus through continued training. In DSA, a lightweight "lightning indexer" scores earlier tokens, and each query then attends only to the highest-scoring ones. Second, a stable RL protocol raises the post-training budget beyond a tenth of the pre-training cost. Third, a pipeline automatically builds many tool-use environments and tasks for agentic RL. The resulting model performs roughly on par with GPT-5 and Kimi-k2-thinking on reasoning benchmarks while costing much less to serve on long contexts. A variant with relaxed length limits, DeepSeek-V3.2-Speciale, matches Gemini-3.0-Pro and reaches gold-medal level at several 2025 olympiads. The work matters because it narrows the open–closed gap and releases both an efficient architecture and an agent-data recipe. The model still trails on world knowledge and token efficiency.

Key points

  • DSA design: core attention cost falls from O(L²) to O(Lk), with 2048 key-value tokens selected per query. The indexer uses ReLU, a small number of heads and FP8. DSA is implemented on top of MLA in MQA mode.
  • Two-stage continued pre-training at 128K context:
    • Indexer warm-up with dense attention and a KL-divergence objective: 1000 steps, about 2.1B tokens.
    • Sparse training of the full model: 15000 steps, about 943.7B tokens.
  • Parity with the predecessor: DeepSeek-V3.2-Exp shows no regression against V3.1-Terminus on standard benchmarks, ChatbotArena Elo, AA-LCR (+4 points in reasoning mode) and Fiction.liveBench.
  • Agentic synthesis: the pipeline produced more than 1,800 environments and 85,000 prompts. The tasks are hard: DeepSeek-V3.2-Exp scores 12% Pass@1 on a sample of them, and GPT-5-Thinking scores 62%. RL on this synthetic data alone improved Tau2Bench, MCP-Mark and MCP-Universe results.
  • Results:
    • DeepSeek-V3.2-Speciale reaches gold-medal level at IMO 2025, IOI 2025, ICPC World Final 2025 and CMO 2025.
    • On BrowseComp, the simple "Discard-all" context-management strategy scores 67.6. This is comparable to parallel sampling while using far fewer steps.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Cited by

314 papers of our library cite it (of the 140673 whose bibliographies we read); 1 of them have a page here, 313 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2512.02556 · arXiv 2512.02556

@misc{deepseekai2025deepseek,
  title = {{DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models}},
  author = {DeepSeek-AI and Liu, Aixin and Mei, Aoxue and Lin, Bangcai and Xue, Bing and Wang, Bingxuan and Xu, Bingzheng and Wu, Bochao and Zhang, Bowei and Lin, Chaofan and Dong, Chen and Lu, Chengda and Zhao, Chenggang and Deng, Chengqi and Xu, Chenhao and Ruan, Chong and Dai, Damai and Guo, Daya and Yang, Dejian and Chen, Deli and Li, Erhang and Zhou, Fangqi and Lin, Fangyun and Dai, Fucong and Hao, Guangbo and Chen, Guanting and Li, Guowei and Zhang, H. and Xu, Hanwei and Li, Hao and Liang, Haofen and Wei, Haoran and Zhang, Haowei and Luo, Haowen and Ji, Haozhe and Ding, Honghui and Tang, Hongxuan and Cao, Huanqi and Gao, Huazuo and Qu, Hui and Zeng, Hui and Huang, Jialiang and Li, Jiashi and Xu, Jiaxin and Hu, Jiewen and Chen, Jingchang and Xiang, Jingting and Yuan, Jingyang and Cheng, Jingyuan and Zhu, Jinhua and Ran, Jun and Jiang, Junguang and Qiu, Junjie and Li, Junlong and Song, Junxiao and Dong, Kai and Gao, Kaige and Guan, Kang and Huang, Kexin and Zhou, Kexing and Huang, Kezhao and Yu, Kuai and Wang, Lean and Zhang, Lecong and Wang, Lei and Zhao, Liang and Yin, Liangsheng and Guo, Lihua and Luo, Lingxiao and Ma, Linwang and Wang, Litong and Zhang, Liyue and Di, M.S. and Xu, M.Y and Zhang, Mingchuan and Zhang, Minghua and Tang, Minghui and Zhou, Mingxu and Huang, Panpan and Cong, Peixin and Wang, Peiyi and Wang, Qiancheng and Zhu, Qihao and Li, Qingyang and Chen, Qinyu and Du, Qiushi and Xu, Ruiling and Ge, Ruiqi and Zhang, Ruisong and Pan, Ruizhe and Wang, Runji and Yin, Runqiu and Xu, Runxin and Shen, Ruomeng and Zhang, Ruoyu and Liu, S.H. and Lu, Shanghao and Zhou, Shangyan and Chen, Shanhuang and Cai, Shaofei and Chen, Shaoyuan and Hu, Shengding and Liu, Shengyu and Hu, Shiqiang and Ma, Shirong and Wang, Shiyu and Yu, Shuiping and Zhou, Shunfeng and Pan, Shuting and Zhou, Songyang and Ni, Tao and Yun, Tao and Pei, Tian and Ye, Tian and Yue, Tianyuan and Zeng, Wangding and Liu, Wen and Liang, Wenfeng and Pang, Wenjie and Luo, Wenjing and Gao, Wenjun and Zhang, Wentao and Gao, Xi and Wang, Xiangwen and Bi, Xiao and Liu, Xiaodong and Wang, Xiaohan and Chen, Xiaokang and Zhang, Xiaokang and Nie, Xiaotao and Cheng, Xin and Liu, Xin and Xie, Xin and Liu, Xingchao and Yu, Xingkai and Li, Xingyou and Yang, Xinyu and Li, Xinyuan and Chen, Xu and Su, Xuecheng and Pan, Xuehai and Lin, Xuheng and Fu, Xuwei and Wang, Y.Q. and Zhang, Yang and Xu, Yanhong and Ma, Yanru and Li, Yao and Li, Yao and Zhao, Yao and Sun, Yaofeng and Wang, Yaohui and Qian, Yi and Yu, Yi and Zhang, Yichao and Ding, Yifan and Shi, Yifan and Xiong, Yiliang and He, Ying and Zhou, Ying and Zhong, Yinmin and Piao, Yishi and Wang, Yisong and Chen, Yixiao and Tan, Yixuan and Wei, Yixuan and Ma, Yiyang and Liu, Yiyuan and Yang, Yonglun and Guo, Yongqiang and Wu, Yongtong and Wu, Yu and Cheng, Yuan and Ou, Yuan and Xu, Yuanfan and Wang, Yuduan and Gong, Yue and Wu, Yuhan and Zou, Yuheng and Li, Yukun and Xiong, Yunfan and Luo, Yuxiang and You, Yuxiang and Liu, Yuxuan and Zhou, Yuyang and Wu, Z.F. and Ren, Z.Z. and Zhao, Zehua and Ren, Zehui and Sha, Zhangli and Fu, Zhe and Xu, Zhean and Xie, Zhenda and Zhang, Zhengyan and Hao, Zhewen and Gou, Zhibin and Ma, Zhicheng and Yan, Zhigang and Shao, Zhihong and Huang, Zhixian and Wu, Zhiyu and Li, Zhuoshu and Zhang, Zhuping and Xu, Zian and Wang, Zihao and Gu, Zihui and Zhu, Zijia and Li, Zilin and Zhang, Zipeng and Xie, Ziwei and Gao, Ziyi and Pan, Zizheng and Yao, Zongqing and Feng, Bei and Li, Hui and Cai, J.L. and Ni, Jiaqi and Xu, Lei and Li, Meng and Tian, Ning and Chen, R.J. and Jin, R.L. and Li, S.S. and Zhou, Shuang and Sun, Tianyu and Li, X.Q. and Jin, Xiangyue and Shen, Xiaojin and Chen, Xiaosha and Song, Xinnan and Zhou, Xinyi and Zhu, Y.X. and Huang, Yanping and Li, Yaohui and Zheng, Yi and Zhu, Yuchen and Ma, Yunxian and Huang, Zhen and Xu, Zhipeng and Zhang, Zhongyu and Ji, Dongjie and Liang, Jian and Guo, Jianzhong and Chen, Jin and Xia, Leyi and Wang, Miaojun and Li, Mingming and Zhang, Peng and Chen, Ruyi and Sun, Shangmian and Wu, Shaoqing and Ye, Shengfeng and T.Wang and Xiao, W.L. and An, Wei and Wang, Xianzu and Sun, Xiaowen and Wang, Xiaoxiang and Tang, Ying and Zha, Yukun and Zhang, Zekai and Ju, Zhe and Zhang, Zhen and Qu, Zihua},
  year = {2025},
  doi = {10.48550/arxiv.2512.02556},
  eprint = {2512.02556},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2512.02556}
}

References

32 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Anthropic. System card: Claude opus 4.5, 2025a. URL https://assets.anthropic.com/m /64823ba7485345a7/Claude-Opus-4-5-System-Card.pdf . ↗
  2. Anthropic. Introducing claude sonnet 4.5, 2025b. URL https://www.anthropic.com/news /claude-sonnet-4-5l . ↗
  3. 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. find
  4. V. Barres, H. Dong, S. Ray, X. Si, and K. Narasimhan. 𝜏 2 -bench: Evaluating conversational agents in a dual-control environment, 2025. URL https://arxiv.org/abs/2506.07982 . ↗
  5. DeepMind. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025a. find our card
  6. G. DeepMind. Gemini 3 pro model card, 2025b. URL https://storage.googleapis.com /deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf . ↗
  7. 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 . find
  8. DeepSeek-AI. Deepseek-v3 technical report, 2024. URL https://arxiv.org/abs/2412.194 37 . ↗
  9. DeepSeek-AI. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081):633–638, 2025. find
  10. EvalSys. Mcpmark leaderboard, 2025. URL https://mcpmark.ai/leaderboard . ↗
  11. 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. find
  12. 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. — [arXiv:2508.14704](https://arxiv.org/abs/2508.14704) find
  13. 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… find
  14. MiniMax. https://www.minimax.io/news/minimax-m2, 2025. URL https://www.minimax. io/news/minimax-m2 . ↗
  15. MoonShot. Introducing kimi k2 thinking, 2025. URL https://moonshotai.github.io/Kim i-K2/thinking.html . ↗
  16. OpenAI. Learning to reason with llms, 2024a. URL https://openai.com/index/learnin g-to-reason-with-llms/ . ↗
  17. 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/ . ↗
  18. OpenAI. Introducing gpt-5, 2025. URL https://openai.com/index/introducing-gpt-5 / . ↗
  19. 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. find our card
  20. Qwen. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388 . ↗ our card
  21. 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. find our card
  22. J. Schulman. Approximating KL divergence, 2020. URL http://joschu.net/blog/kl-app rox.html . ↗
  23. Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, M. Zhang, Y. K. Li, Y. Wu, and D. Guo. Deepseekmath: 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 . find our card
  24. 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. find
  25. N. Shazeer. Fast transformer decoding: One write-head is all you need. CoRR, abs/1911.02150, 2019. URL http://arxiv.org/abs/1911.02150 . ↗
  26. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. pages 5998–6008, 2017. URL https://proceedings.neur ips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html . ↗
  27. 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… ↗ our card
  28. 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. find
  29. 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 . — [arXiv:2504.21798](https://arxiv.org/abs/2504.21798) find
  30. 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. Pilehvar, editors, Proceedings of t… find our card
  31. ZhiPu-AI. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models. arXiv preprint arXiv:2508.06471, 2025. find our card
  32. 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. — [arXiv:2504.19314](https://arxiv.org/abs/2504.19314) find