1、DeepSeek-V3.2:Pushing the Frontier of OpenLarge Language ModelsDeepSeek-AIAbstractWe 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 asfollows:(1)DeepSeek Sparse Attentio
2、n(DSA):We introduce DSA,an efficient attention mecha-nism that substantially reduces computational complexity while preserving model performancein long-context scenarios.(2)Scalable Reinforcement Learning Framework:By implementinga robust reinforcement learning protocol and scaling post-training com
3、pute,DeepSeek-V3.2performs 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,achievinggold-medal performance in both the 2025 International Mathematical Olympiad(IMO)and theInternational Olympiad
4、in Informatics(IOI).(3)Large-Scale Agentic Task Synthesis Pipeline:To integrate reasoning into tool-use scenarios,we developed a novel synthesis pipeline thatsystematically generates training data at scale.This methodology facilitates scalable agenticpost-training,yielding substantial improvements i
5、n generalization and instruction-followingrobustness within complex,interactive environments.AIME 2025(Pass1)HMMT 2025(Pass1)HLE(Pass1)Codeforces(Rating)SWEVerified(Resolved)TerminalBench 2.0(Acc)2Bench(Pass1)ToolDecathlon(Pass1)020406080100Accuracy/Pass1(%)96.093.194.687.095.099.290.288.379.297.530
6、.625.126.313.737.72701238625371480270873.174.977.276.246.435.242.854.280.380.284.785.435.229.038.636.4Reasoning CapabilitiesAgentic Capabilities050010001500200025003000Codeforces RatingDeepSeek-V3.2-SpecialeDeepSeek-V3.2-ThinkingGPT-5-HighClaude-4.5-SonnetGemini-3.0-ProFigure 1|Benchmark of DeepSeek
0731-84720580
商务合作：really158d
研搜
微信扫码登录
手机快捷登录
账号登录
