DeepSeek Math-V2 Hits IMO Gold and Goes Open Source
DeepSeek, a startup out of Hangzhou, China, put out Math-V2, the first open AI model to score at gold-medal level on the International Mathematical Olympiad. It handles problems that demand real creativity and step-by-step reasoning, stuff that’s tripped up other AIs before. You can grab it free on Hugging Face or GitHub right now.
The Scores That Matter
Math-V2 nailed five out of six problems from IMO 2024, enough for gold—only about 8% of human contestants pull that off each year, as WebProNews points out. It also hit gold on the 2024 Chinese Mathematical Olympiad, per Digital Watch Observatory and Moneycontrol.
On top of that, it scored 118 out of 120 on the 2024 Putnam exam, beating the best human marks there. Google DeepMind and OpenAI got similar IMO results earlier this year, but they kept their models closed, per SQ Magazine.
- IMO 2024: 5/6 problems solved (WebProNews)
- Putnam 2024: 118/120 (WebProNews)
- CMO 2024: Gold level (Digital Watch Observatory, Moneycontrol)
What’s Under the Hood
The model uses a mixture-of-experts setup, which picks the right sub-parts of the network for math tasks without burning through crazy compute. It leans on self-verification—basically, it double-checks its own work by trying different paths to a solution. That cut down errors in theorem proving and tough puzzles, according to coverage from WebProNews, Interesting Engineering, and AOL.
Why Open Source Changes Things
Under Apache 2.0, anyone can tweak it, run it locally, or build on top—no paywalls. Hugging Face CEO Clement Delangue called it like getting “the brain of one of the best mathematicians in the world for free,” as Moneycontrol reported. While US companies lock down their top math AIs, this drops the gate for researchers everywhere. Digital Watch Observatory notes it could speed up work in science and engineering since devs don’t need elite hardware.
DeepSeek built this fast after closed models from the big players hit the same marks in July. It’s efficient too—runs cheaper than rivals. Folks on X are already fine-tuning it for tutoring or simulations.
