DeepSeek V3.2

DeepSeek officially launched DeepSeek-V3.2 and a high-compute variant DeepSeek-V3.2-Speciale on December 1, 2025, with V3.2 designated the official successor to V3.2-Exp and available across App, Web, and API. The release page highlights "Thinking in Tool-Use," described as DeepSeek's first model to integrate thinking directly into tool-use, supporting tool-use in both thinking and non-thinking modes. Open-source weights for both V3.2 and V3.2-Speciale were published on Hugging Face alongside the technical report. The announcement emphasizes reasoning-first agent capability: V3.2 is positioned as a daily driver at GPT-5-level balanced performance, while V3.2-Speciale is reported to rival Gemini-3.0-Pro and attain gold-medal results at IMO, CMO, ICPC World Finals, and IOI 2025. A new large-scale agent training data synthesis method covers 1,800+ environments with 85k+ complex instructions, and the V3.2 API retains the same usage pattern as V3.2-Exp. V3.2-Speciale is API-only, no tool calls, with the temporary endpoint expiring December 15, 2025.

The official Hugging Face model card reiterates DeepSeek-V3.2's three pillars: DeepSeek Sparse Attention (DSA) for long-context efficiency, a scalable reinforcement learning framework producing GPT-5-class reasoning with V3.2-Speciale on par with Gemini-3.0-Pro, and a large-scale agentic task synthesis pipeline that yields gold-medal IMO and IOI 2025 results. The card notes that final submissions for IOI 2025, ICPC World Finals, IMO 2025, and CMO 2025 are released under assets/olympiad cases so the community can perform secondary verification of the reasoning claims. Developer-facing changes include a revised chat template with a new tool-calling format and an introduced "thinking with tools" capability, supported by Python encoding scripts and test cases showing how to convert OpenAI-compatible messages into V3.2 input strings and parse outputs. The model card confirms the sibling open-source release of V3.2-Speciale weights and links the technical report PDF. These template updates directly affect downstream integrations that consume DeepSeek-V3.2 chat completions.

The arXiv technical report introduces DeepSeek-V3.2 and details three breakthroughs: DeepSeek Sparse Attention (DSA), an efficient attention mechanism reducing computational complexity while preserving long-context performance; a scalable reinforcement learning protocol that yields GPT-5-comparable results; and a large-scale agentic task synthesis pipeline that systematically generates training data at scale to integrate reasoning into tool-use scenarios. The high-compute variant V3.2-Speciale reportedly surpasses GPT-5 and matches Gemini-3.0-Pro, with reported gold-medal results in both IMO 2025 and IOI 2025. Figure 1 of the paper benchmarks DeepSeek-V3.2 against proprietary and open counterparts, with HMMT 2025 reported using the February competition to stay consistent with baselines and HLE using the text-only subset. Authors frame the release within a divergence between accelerating closed-source frontier models and continued open-source progress from labs including Qwen, ZhiPu, MiniMax, and MoonShot. The report establishes V3.2 as a reasoning-and-agent-focused open-weights flagship with explicit cost-efficiency design choices.