DeepSeek releases V3.2 and gold-medal-level V3.2-Speciale

Background

DeepSeek-V3.2 was released on 1 December 2025 as the official successor to the experimental DeepSeek-V3.2-Exp ^[1]. Its positioning went beyond experiments: DeepSeek presented it as a model for everyday use, as the quotation in the next section shows ^[1].

What happened

DeepSeek released V3.2 on 1 December 2025 as the official successor to DeepSeek-V3.2-Exp. The model was live the same day in DeepSeek's app, on its web chat and through its API ^[1]. The same release included V3.2-Speciale, a companion model designed to maximize reasoning ^[1][2].

DeepSeek called V3.2 'your daily driver at GPT-5 level performance', a phrase that signals the company's intent to position it as a model for everyday work ^[1].

Key details

DeepSeek described V3.2 as its first model to integrate thinking directly into tool use. It supports tool calls in both thinking and non-thinking modes, so a model can call external functions while it is reasoning or without reasoning first ^[1].

The company also reported a new synthesis method for agent training that covered more than 1,800 environments and 85,000 complex instructions ^[1]. Reported scores included 82.4 on GPQA Diamond, 85.0 on MMLU-Pro, 70.0 on SWE-bench Verified, 70.2 on SWE-bench Multilingual and 39.6 on Terminal-Bench 2.0 ^[2].

V3.2-Speciale was described as rivaling Gemini-3.0-Pro. DeepSeek said it reached gold-level results at IMO 2025, CMO 2025, the ICPC World Finals and IOI 2025 ^[1][2]. Its limits were stated clearly. It needs more tokens, does not support tool calls, and was served only through a temporary API endpoint until 15 December 2025, at the same price as V3.2 ^[1].

Both models' weights were published on Hugging Face along with a technical report ^[1].

Impact

With the V3.2 release, DeepSeek's open-weight line moved from experimental releases to a model that could be used in the company's own app and API from day one. The Speciale version also showed the company's interest in competition-level reasoning, with competition submissions published for independent checking ^[2].