DeepSeek V3.2 Exp

DeepSeek released V3.2 Exp, a 685B-parameter mixture-of-experts model with 37B active parameters per token, featuring DeepSeek Sparse Attention (DSA) for fine-grained sparse attention. The model, built on V3.1-Terminus and released under MIT license, was later superseded by V4 within about seven months.

DeepSeek V3.2 Exp MoE workstation685B total, ~37B active per token MoE, 256 routed experts - shares the V3 backbone. Adds DeepSeek Sparse Attention DSA : a lightning indexer FP8, Hadamard dot-product selects top-2048 tokens per query, then MLA runs only over those - the first fine-grained sparse attention from DeepSeek and the precursor to V4's CSA. Built on V3.1-Terminus. Experimental; superseded by V4 within ~7 months. Open weights under MIT. coding reasoning - 685.0B - 128k - mit - Sep 2025 Scores Coding 80 Reasoning 86 General 83 PRICE HISTORY Inference cost over time Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API. Loading price history...