Dec 02, 2025 11:34:00

'DeepSeek-V3.2' is now available, an open model with performance equivalent to GPT-5 and Gemini 3, and is available for free

DeepSeek, a Chinese AI development company, released the inference models ' DeepSeek-V3.2 ' and ' DeepSeek-V3.2-Speciale ' on December 1, 2025. Both models are open models with publicly available model data, yet they have achieved benchmark scores comparable to GPT-5 and Gemini 3.

deepseek-ai/DeepSeek-V3.2 · Hugging Face

https://huggingface.co/deepseek-ai/DeepSeek-V3.2

deepseek-ai/DeepSeek-V3.2-Speciale · Hugging Face
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale

DeepSeek-V3.2 is the official release of the experimental model ' DeepSeek-V3.2-Exp ,' which was released in 2025, and is positioned as a model with performance equivalent to GPT-5. DeepSeek-V3.2-Speciale is a model with enhanced computational capabilities, and is capable of winning gold medals at the International Mathematical Olympiad and the International Olympiad in Informatics.

Below is a table listing various benchmark scores for 'GPT-5 High,' 'Gemini 3.0 Pro,' 'Kimi-K2 Thinking,' 'DeepSeek-V3.2,' and 'DeepSeek-V3.2-Speciale.' DeepSeek-V3.2 recorded scores equivalent to GPT-5 High, while DeepSeek-V3.2-Speciale exceeded the score of GPT-5 High. DeepSeek-V3.2-Speciale also surpassed the score of Gemini 3.0 Pro in many tests.

DeepSeek-V3.2 can also run agent tasks using tools, where it outperformed open models such as Kimi-K2 Thinking and MiniMax M2, and achieved scores close to those of GPT-5 High and Gemini 3.0 Pro.

DeepSeek-V3.2 is also characterized by its high computational efficiency. The graph below shows the relationship between the increase in tokens processed and the cost. In DeepSeek-V3.1-Terminus (blue), the cost increased as the number of processed tokens increased, but in DeepSeek-V3.2 (orange), the increase in cost was suppressed.

The model data for DeepSeek-V3.2 and DeepSeek-V3.2-Speciale is available on Hugging Face and is also available via API .