DeepSeek V3.2: Sparse Attention and Reinforcement Learning Updates

Summary

Sebastian Raschka’s technical overview traces DeepSeek from V3 and R1 through V3.1, V3.2-Exp, DeepSeekMath V2, and the December 1, 2025 V3.2 release. V3 and R1 share a Mixture-of-Experts architecture and Multi-Head Latent Attention (MLA), while R1 added reinforcement learning with verifiable rewards (RLVR) and GRPO to improve reasoning. V3.1 moved to a hybrid model that lets users switch between instruct and reasoning modes. V3.2-Exp introduced DeepSeek Sparse Attention (DSA), combining a learned lightning indexer with a token selector that retains only high-scoring past tokens. Together with MLA, this changes attention cost from quadratic in sequence length to approximately O(Lk), where k is the selected-token count, although the mechanism can reduce model quality and requires custom inference support. DeepSeekMath V2 extended RLVR with an LLM proof verifier, a meta-verifier, and self-refinement to address the problem that correct final answers do not necessarily imply correct reasoning. The article reports that the verifier’s analysis quality rose from 0.85 to 0.96 in the described setup, and that up to eight refinement iterations improved accuracy without clear saturation. DeepSeek V3.2 keeps the V3.2-Exp architecture and combines sparse attention with revised training. Its rewards include rule-based outcome rewards, length penalties, language consistency rewards, and a generative reward model for general tasks, making the pipeline hybrid rather than purely verifier-based. GRPO changes include domain-specific KL strengths, an importance-weighted KL estimate, off-policy sequence filtering, preserved MoE routing and sampling masks, while retaining the original advantage normalization. V3.2-Speciale is trained only on reasoning data with a reduced length penalty to support longer, more accurate responses at higher inference cost. The article also discusses mHC, a separate DeepSeek proposal that constrains widened residual connections to improve training stability and convergence with modest overhead.