DeepSeek has released a powerful open-source AI model, Math-V2, capable of solving International Mathematical Olympiad problems at gold medal level.
Quick Summary – TLDR:
- DeepSeek’s Math-V2 AI model achieves gold medal-level scores at the prestigious International Mathematical Olympiad (IMO)
- The model is fully open-source and available on Hugging Face and GitHub
- It matches or exceeds the performance of Google’s Gemini and OpenAI’s GPT models
- DeepSeek’s approach includes self-verifiable reasoning, a key innovation for AI in advanced mathematics
What Happened?
Chinese AI startup DeepSeek has released Math-V2, an AI model that performs at gold medal level in the International Mathematical Olympiad (IMO), a global competition known for its intensely challenging math problems. Unlike models from OpenAI and Google DeepMind, DeepSeek has made its system freely available, promoting openness in AI development.
🚨 DeepSeek just did something wild.
— Robert Youssef (@rryssf_) November 29, 2025
They built a math model that doesn’t just solve problems, it checks its own proofs, criticizes itself, fixes the logic, and tries again until it can’t find a single flaw.
That final part is the breakthrough a model that can verify its own… pic.twitter.com/QE28hxWQU9
Math-V2: A Gold-Level Model Goes Open
The International Mathematical Olympiad is considered the most rigorous mathematics competition in the world. Only around 8 percent of human participants score at the gold medal level. DeepSeek’s Math-V2 AI model has now joined this elite group, solving five out of six problems from the 2024 IMO and also delivering strong results in the Chinese Mathematical Olympiad.
Math-V2 was published on Hugging Face and GitHub under a permissive license, allowing developers to modify, repurpose, and run the model locally. This is a major contrast to similar models from OpenAI and Google, which are restricted to premium or closed access.
Clément Delangue, CEO of Hugging Face said:
DeepSeek vs GPT and Gemini: A Performance Comparison
Earlier this year, both OpenAI’s GPT and Google’s Gemini announced that they had reached gold-level performance on IMO problems. However, these models remain inaccessible to the public. DeepSeek not only matched their performance but also outperformed them on some benchmarks:
- On Google DeepMind’s IMO-ProofBench, a benchmark designed to test deep mathematical reasoning:
- DeepSeek’s Math-V2 scored 99% on the baseline test.
- Gemini DeepSeek scored 89%.
- GPT-5 scored 59%.
- On the advanced test, Math-V2 scored 61.9%, just behind Gemini DeepSeek’s 63.7%.
These scores highlight DeepSeek’s competitive edge in mathematical reasoning and its commitment to open AI development.
Self-Verification: A Step Toward Smarter AI
One of the standout features of Math-V2 is its ability to self-verify its solutions. Most AI systems improve only when trained on problems with known answers. DeepSeek’s model goes further by evaluating the validity and consistency of its own reasoning, even for problems with no existing solution. This allows it to tackle more open-ended and complex challenges.
According to DeepSeek researchers, this self-checking method is a key to building AI that genuinely understands mathematical concepts, not just memorizes solutions. It marks a shift from solving simple benchmarks to handling real-world, high-level problems.
Global Impact and Democratization of AI
DeepSeek’s open-source move could have a transformative impact on both education and research in AI and mathematics. By making its high-performing model freely accessible, it helps reduce the entry barrier for developers and researchers around the world.
While U.S. tech giants like OpenAI, Google, and Anthropic focus on monetization through subscriptions and restricted access, Chinese AI firms are leveraging open access as a strategic differentiator, especially amid restrictions on advanced hardware like NVIDIA AI chips.
SQ Magazine’s Takeaway
I think this move by DeepSeek is more than just another AI milestone. It’s a statement. In a world where top AI breakthroughs are usually locked behind paywalls, DeepSeek is showing that elite-level AI can be shared, not just sold. This kind of openness doesn’t just fuel innovation, it empowers researchers, educators, and even hobbyists around the world to participate in shaping the future of AI. And let’s be honest, a math model that can reason, verify itself, and beat the best in the world? That’s not just smart. That’s a game-changer.
