By ByteBot
DeepSeek just did what OpenAI and Google wouldn’t: give away a gold-medal AI for free. The Chinese AI startup released DeepSeekMath-V2 on November 28, 2025, the first open-source model to achieve International Mathematical Olympiad (IMO) gold medal level performance. It solved 5 out of 6 IMO 2025 problems and scored 118 out of 120 on the Putnam 2024 exam, crushing the highest human score of 90.
While OpenAI and Google DeepMind achieved identical IMO scores this year, they kept their models locked behind proprietary APIs. DeepSeek released theirs under an MIT license on Hugging Face and GitHub, letting any developer access world-class mathematical reasoning AI without API fees or vendor lock-in.
DeepSeek Math V2 Matches OpenAI and Google Performance
DeepSeekMath-V2’s IMO performance matches OpenAI’s o1 and Google’s Gemini models exactly. All three solved the same 5 problems at IMO 2025, earning gold medal scores of 35 out of 42 points. However, the difference is striking: DeepSeek made its 685-billion-parameter model freely available under the MIT license, while its competitors charge enterprise customers millions for API access.
Moreover, the Putnam exam results tell an even more impressive story. DeepSeek’s 118 out of 120 score didn’t just beat other AI models—it destroyed the highest human score of 90. For context, the Putnam exam is notoriously brutal: 3,988 students took it in 2024, with a median score of 2 and an average of 8. DeepSeek solved 11 of 12 problems completely.
There’s a caveat worth noting: the AI used “scaled test-time compute,” meaning it had more computational resources than the six-hour limit humans face. Still, surpassing the best mathematical minds by 28 points is remarkable.
Self-Verification Makes DeepSeek Math More Reliable
DeepSeek’s technical innovation goes beyond raw performance. The model uses a novel “verifier-generator” dual architecture that mimics how human mathematicians check their work. A generator proposes step-by-step solutions, while a verifier examines each step’s logic, marking it “valid,” “incomplete,” or “incorrect.” If problems arise, the generator revises and fixes mistakes.
Furthermore, this addresses a critical AI reliability issue: models often reach correct answers through flawed reasoning. DeepSeek’s approach ensures sound logic throughout the proof, making it more trustworthy for real-world mathematical applications like scientific computing and theorem proving.
Open Source AI Becomes Competitive Advantage
DeepSeek’s release intensifies the debate over whether cutting-edge AI should be open or closed. Proprietary advocates argue advanced AI is “dual-use technology” like nuclear weapons, requiring controlled access to prevent misuse. In contrast, open-source supporters counter that transparency enables faster vulnerability detection and prevents vendor monopolies—the same argument that won out for cryptography.
Nevertheless, DeepSeek proves open-source can compete at the highest levels. The company trained its model for $6 million, compared to OpenAI’s $100 million for GPT-4, using one-tenth the computing power of Meta’s comparable model. Efficiency is beating brute-force scaling.
The geopolitical dimension is hard to ignore. China’s strategy appears to be giving away top models to dominate through adoption and developer mindshare, while US tech giants charge premium prices and maintain tight control. Consequently, one provocative question emerges: Why is China giving away what American companies charge millions for?
Access DeepSeek Math V2 Now for Free
Developers can download DeepSeekMath-V2 today from Hugging Face or GitHub. The MIT license allows unrestricted commercial use, fine-tuning for specific applications, and deployment without ongoing API costs.
Use cases include automated theorem proving, mathematical problem-solving in research, educational tools with step-by-step explanations, and scientific computing verification. Additionally, if you need mathematical reasoning AI, you can now get gold-medal performance without paying OpenAI’s API fees or accepting Google’s terms.
Will Silicon Valley Respond?
DeepSeek’s move puts pressure on OpenAI and Google to justify their proprietary approaches. If an open-source model can match their performance at a fraction of the cost, why pay for closed access? This could reshape industry power dynamics, with open-source becoming not just the “free alternative” but the preferred choice for developers prioritizing transparency, control, and cost efficiency.
The question now: Will Silicon Valley respond by opening their models, or double down on proprietary AI and hope regulatory moats protect their position?
