DeepSeek releases DeepSeek-V3.2 and V3.2-Speciale

DeepSeek released two open-weight models on the same day: DeepSeek-V3.2, a general-purpose chat model made available across its app, web interface and API, and DeepSeek-V3.2-Speciale, an extended-reasoning variant offered only through a temporary API endpoint, scheduled to expire two weeks later.

DeepSeek pitched V3.2 as “your daily driver at GPT-5 level performance,” and said it was the company’s first model to integrate reasoning directly into tool use, working in both a “thinking” and a faster non-thinking mode. It also introduced what DeepSeek described as a large agent-training data synthesis method spanning more than 1,800 environments and over 85,000 complex instructions, aimed at improving the model’s reliability on multi-step, tool-using tasks. V3.2-Speciale, described as having “maxed-out reasoning capabilities” comparable to Google’s Gemini 3 Pro, was reported to reach gold-medal-level results at the International Mathematical Olympiad, the China Mathematical Olympiad, the ICPC World Finals and the International Olympiad in Informatics — competition-level maths and coding benchmarks that had become a standard proving ground for reasoning models.

Both figures came from DeepSeek’s own release notes rather than independent evaluation, and V3.2-Speciale’s restriction to a short-lived, tool-free API endpoint suggested it was released as a research or evaluation preview rather than a production model. Weights for both were published on Hugging Face.

The release continued a pattern set by DeepSeek’s R1 release in January 2025: a Chinese lab publishing open weights claimed to be competitive with the leading closed US models within weeks of their release, keeping pressure on the argument that frontier capability required a closed, proprietary model.