DeepSeek V3.2

2025-12-01 · Publisher : DeepSeek · Domain : Artificial Intelligence

DeepSeek releases V3.2 and V3.2-Speciale: the former reaches GPT-5 level according to the vendor, the latter rivals Gemini 3 Pro and wins gold at the IMO and IOI 2025.

### Context V3.2-Exp had validated DSA sparse attention in September. V3.2 is its official successor, released a few days after Gemini 3 Pro (November 2025) and after GPT-5 (August 2025), which were then the closed references. The goal is to show that an open model can hold its own against these models in reasoning and tool use. ### What the model brings - An “everyday” model that balances reasoning length and quality, presented at GPT-5 level by the vendor. - V3.2-Speciale, a maximum-reasoning variant, announced at Gemini 3.0 Pro level, with gold-level results at the IMO, CMO, ICPC World Finals and IOI 2025. - Thinking directly integrated into tool calls, usable in thinking mode as well as without thinking. - Large-scale agent data synthesis: more than 1,800 environments and more than 85,000 complex instructions. - A scalable reinforcement learning framework and DSA carried over from V3.2-Exp. - Publication of the final olympiad submissions for community verification. ### Architecture and specifications - Type: Mixture-of-Experts, 685 billion total parameters (Hugging Face), same structure as V3.2-Exp. - Context: not included in the sources consulted. Data, cutoff: not disclosed. - License: MIT (code and weights), commercial use allowed. - Format novelty: no Jinja template, but a Python encoding folder; new `developer` role reserved for search agents. - Speciale: reasoning only, no tool calls. - Weights: huggingface.co/deepseek-ai/DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. Recommended settings: temperature 1.0, top_p 0.95. ### Cost and access | Item | Detail | |---|---| | V3.2 (API) | Same usage as V3.2-Exp; V3.2-Exp pricing: 0.028 / 0.28 / 0.42 USD per million tokens (cache hit / cache miss / output) | | V3.2-Speciale (API) | Temporary endpoint until December 15, 2025, 15:59 UTC, “same pricing as V3.2”, without tool calls | | Hardware | Not disclosed (see V3.2-Exp repository) | ### Benchmarks | Benchmark | GPT-5 High | Gemini-3.0 Pro | Kimi-K2 Thinking | V3.2 Thinking | V3.2 Speciale | |---|---|---|---|---|---| | AIME 2025 | 94.6 | 95.0 | 94.5 | 93.1 | 96.0 | | HMMT Feb. 2025 | 88.3 | 97.5 | 89.4 | 92.5 | 99.2 | | IMOAnswerBench | 76.0 | 83.3 | 78.6 | 78.3 | 84.5 | | LiveCodeBench | 84.5 | 90.7 | 82.6 | 83.3 | 88.7 | | Codeforces (rating) | 2537 | 2708 | not published | 2386 | 2701 | | GPQA Diamond | 85.7 | 91.9 | 84.5 | 82.4 | 85.7 | | Humanity's Last Exam | 26.3 | 37.7 | 23.9 | 25.1 | 30.6 | Source: table from the official blog post of December 1, 2025, scores measured and published by the vendor. ### Versus closed models At the time of its release, the reference closed models compared by the vendor were GPT-5 High and Gemini 3.0 Pro. V3.2 Thinking is slightly behind GPT-5 High on AIME 2025 (93.1 versus 94.6), LiveCodeBench (83.3 versus 84.5) and GPQA Diamond (82.4 versus 85.7). Speciale exceeds GPT-5 High on AIME 2025 (96.0 versus 94.6), HMMT (99.2 versus 88.3), IMOAnswerBench (84.5 versus 76.0) and Humanity's Last Exam (30.6 versus 26.3). Against Gemini 3.0 Pro, Speciale is ahead on AIME (96.0 versus 95.0) and HMMT (99.2 versus 97.5), behind on Codeforces (2701 versus 2708), GPQA (85.7 versus 91.9) and Humanity's Last Exam (30.6 versus 37.7). Cost is the counterpoint: Speciale consumes more tokens (for example 45k versus 18k for Gemini on IMOAnswerBench, as indicated in the vendor's table). For an open choice, V3.2 offers a level announced as close to GPT-5 with an MIT license. ### Key takeaways V3.2 is the first open model of the period to show, in a vendor table, scores comparable to those of the two major closed models of the moment. The Speciale weights published on Hugging Face are reserved for pure reasoning, without tools.

Tags : LLM, Open Source, DeepSeek, MIT, MoE, Raisonnement, Agents