671B nominal total / 37B nominal active per token (each MoE); 671.878B main-model tensors plus 13.477B packaged MTP tensors
128K tokens supported; 163,840 configured positional ceiling
MIT
Two text-only post-trained checkpoints sharing a 61-layer DeepSeekMoE/MLA architecture with 256 routed experts, eight selected per token and one shared expert. DeepSeek Sparse Attention uses a 64-head, 128-dimensional lightning indexer selecting up to 2,048 prior tokens. Released weights mix FP8 E4M3 and BF16, with UE8M0-format block scaling. The shared sparse base underwent 2.1B-token indexer warm-up and 943.7B-token sparse continued pretraining; these are inherited base-training stages, not an additional December run. Standard V3.2 uses specialist-generated thinking/non-thinking distillation data followed by thousands of mixed GRPO steps for reasoning, agents and human alignment, with length and language-consistency rewards. Agent synthesis includes 1,827 general-agent environments and over 85,000 prompts across four agent domains. Speciale instead trains exclusively on reasoning data with reduced length penalties and DeepSeekMath-V2 proof data/rewards; it consumes more tokens and does not support tool calling. Standard V3.2 is DeepSeek's first model to combine thinking with tools, while also supporting non-thinking tool use. Developer-reported pass@1 is 93.1% AIME 2025 and 92.5% HMMT February for standard, versus 96.0% and 99.2% for Speciale. Standard achieves 73.1% SWE-bench Verified with an internal agent framework and 46.4% Terminal Bench 2.0 with Claude Code; BrowseComp rises from 51.4% to 67.6% with context management. Speciale reaches gold-level competition scores (IMO 35/42, CMO 102/126, IOI 492/600, ICPC 10/12) using filtered multi-candidate submissions or generate-verify-refine workflows, not standalone one-shot medals. The report identifies knowledge breadth, token efficiency and long-trajectory context overflow as remaining weaknesses. Launch packaging supplies Python message encoding rather than a Jinja chat template; its output parser needs production error handling.