DeepSeek V3.2

DeepSeek-V3.2 is DeepSeek's December 2025 flagship, evolved from the experimental V3.2-Exp that introduced DeepSeek Sparse Attention (DSA) — a "lightning indexer" architecture that cuts long-context inference cost by more than half while holding capability steady (MMLU-Pro 85.0, AIME 2025 89.3 vs 88.4 prior).

The weights are MIT-licensed and fully open on the official deepseek-ai Hugging Face org (685B params, FP8), with day-one support in SGLang and vLLM, open-source GPU kernels (TileLang, FlashMLA), and Docker images for NVIDIA H200s / AMD MI350s. DeepSeek pairs this with its own API and an open agentic-tool protocol, and the final V3.2 was reported at 73.1% SWE-Verified and 46.4% Terminal Bench 2.0 (DeepSeek-reported).

For developers, this is the highest-value open frontier model of late 2025: self-hostable with permissive licensing, or cheap via the API.

Provider-reported (DeepSeek, from HF model card / launch charts):

• AIME 2025: 89.3 | MMLU-Pro: 85.0 | Codeforces rating: 2121
• SWE-Verified (final V3.2): 73.1% | SWE Multilingual: 70.2% | Terminal Bench 2.0: 46.4%
• BrowseComp: 40.1 | SimpleQA: 97.1

Independent coverage: VentureBeat (2025-09-30) independently covered the launch, verified the MIT license and the >50% price cut, and noted the efficiency-vs-capability tradeoff. LMSYS SGLang team and vLLM shipped day-one support, implicitly validating the architecture's deployability. No independent end-to-end reproduction of DeepSeek's benchmark numbers was found.