On December 1, 2025, Chinese artificial intelligence company DeepSeek launched the world’s first open-source AI model capable of achieving a gold medal-level score at the annual International Mathematical Olympiad (IMO).
It also unveiled two new versions of an experimental artificial-intelligence model DeepSeek-V3.2 and DeepSeek-V3.2-Speciale.
DeepSeek V3.2 is said to rival OpenAI’s flagship GPT-5 across various reasoning benchmarks, indicating that China’s open-source models continue to compete with Silicon Valley’s leading proprietary systems on several key metrics.
DeepSeek V3.2-Speciale is designed for advanced mathematical computations and extended reasoning tasks, reportedly achieving performance comparable to Google’s latest Gemini-3 Pro.
Built on top of the DeepSeek-V3.2-Exp-Base model, the Math-V2 model is now publicly available on Hugging Face and GitHub under a permissive licence, allowing developers worldwide to freely adapt and build upon it.
Imagine owning the brain of one of the best mathematicians in the world for free to explore it for research, fine-tune it, optimise it and run it on your own hardware.
Hugging Face co-founder and CEO Clement Delangu
Why it stands out
To reach gold-level performance, AI systems must not only solve problems correctly but also show clear reasoning behind their answers. Only about 8% of human IMO participants typically earn a gold medal.
The IMO, held annually since its inception in 1959, is widely considered the most prestigious mathematics competition globally, including for AI systems, due to the complexity of its problems which demand deep insight, creativity, and rigor.
DeepSeek enhanced the rigor of its model’s mathematical reasoning by incorporating a “self-verification” mechanism. This allows the model to automatically check and validate its answers, even for math problems without known solutions. This approach addresses a key limitation in current AI systems, which tend to improve mainly on tasks with easily verifiable solutions.
DeepSeek said its open model aims to make advanced mathematical reasoning technology more accessible to researchers and developers, following similar achievements by Google DeepMind and OpenAI, whose top-performing IMO models remain proprietary.
Google DeepMind made its proprietary model that won a gold medal at the IMO available exclusively to subscribers of its premium Ultra plan. OpenAI CEO Sam Altman stated that the company’s experimental gold medal-winning model would not be publicly accessible for many months.
The company added that Math-V2 earned gold-level results on tasks from both this year’s IMO and the 2024 Chinese Mathematical Olympiad.To reach gold-level performance, AI systems must not only solve problems correctly but also show clear reasoning behind their answers. Only about 8% of human IMO participants typically earn a gold medal.
Why it matters
The launch is part of a broader race between Chinese and U.S. AI firms over who can produce the most capable frontier systems. DeepSeek’s open-source approach gives it a different appeal from rivals such as Google DeepMind and OpenAI, whose best IMO-performing models remain closed or limited to paid users.
Read more:
Google launches Gemini 3 and Ironwood
Nvidia first to reach $5T market cap
OpenAI launches GPT-5 for all users
