DeepSeek V3.2-Exp

2025-09-29 · Publisher : DeepSeek · Domain : Artificial Intelligence

DeepSeek releases V3.2-Exp, an experimental version that introduces DSA sparse attention to make long context less costly, at quality equivalent to V3.1-Terminus, with API prices cut by more than half.

### Context V3.1-Terminus (September 22, 2025) had just stabilized the V3.1 line. V3.2-Exp is presented as an intermediate step toward the next-generation architecture: the vendor is testing training and inference optimizations for long contexts there. The challenge is the cost of attention, which grows with sequence length. ### What the model brings - DeepSeek Sparse Attention: according to the vendor, the first fine-grained sparse attention, with substantial gains in training and inference on long context, for an output quality that is "virtually identical". - A controlled comparison: the training configuration is aligned with V3.1-Terminus to isolate the effect of sparse attention. - Kernels published as open source: TileLang for readability, CUDA (DeepGEMM, FlashMLA) for performance. - Immediate day-one support in vLLM, and SGLang Docker images (H200, MI350, NPU). - An API price cut of more than 50%, effective immediately. ### Architecture and specifications - Type: Mixture-of-Experts, 685 billion parameters in total on Hugging Face (V3.1-Terminus base, 671 billion parameters in the inference configuration, 256 experts). - Innovation: DSA, sparse selection of the tokens to consult by an indexer. - Context: not included in the model card consulted. Training data, active parameters and cutoff: not disclosed. - License: MIT (code and weights), commercial use allowed. - Weights: huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp. - Fix reported on November 17, 2025: a RoPE error in the indexing module of the demo code, corrected in the inference code. ### Cost and access | API (USD per million tokens) | Before | Since September 29, 2025, 10:00 UTC | |---|---|---| | Input, cache hit | 0.07 | 0.028 | | Input, cache miss | 0.56 | 0.28 | | Output | 1.68 | 0.42 | V3.1-Terminus remained available through a temporary API until October 15, 2025, 15:59 UTC. Self-hosting: the reference commands use 8 GPUs (tp 8); exact memory not disclosed. ### Benchmarks | Benchmark | V3.1-Terminus | V3.2-Exp | |---|---|---| | MMLU-Pro | 85.0 | 85.0 | | GPQA Diamond | 80.7 | 79.9 | | Humanity's Last Exam | 21.7 | 19.8 | | LiveCodeBench | 74.9 | 74.1 | | AIME 2025 | 88.4 | 89.3 | | Codeforces (rating) | 2046 | 2121 | | BrowseComp | 38.5 | 40.1 | | SWE Verified | 68.4 | 67.8 | | Terminal-bench | 36.7 | 37.7 | Source: DeepSeek-V3.2-Exp model card on Hugging Face, scores published by the vendor. No closed model appears in it. ### Versus closed models The source contains no comparison with a closed model, so none can be quantified here. The vendor's message is about efficiency: with scores nearly identical to V3.1-Terminus (differences of less than one point on most lines), the output price drops from 1.68 to 0.42 dollars per million tokens, a fourfold reduction. On the open versus closed axis, the contribution is therefore economic: a long context that is cheaper to serve, including in self-hosting thanks to the published kernels. The vendor itself presents the version as experimental. ### Key takeaways V3.2-Exp is a technical milestone more than a capability one: DSA is the building block reused afterward in V3.2. The price cut of more than 50% accompanies this increased efficiency.

Tags : LLM, Open Source, DeepSeek, MIT, MoE, Attention éparse, Contexte long