DEEPSEEK has launched a powerful new AI model for mathematics called DeepSeek-Math-V2. Unlike a calculator that just gives answers, this model is designed for complex, step-by-step reasoning, much like a mathematician working on a difficult proof. The model has reached a gold medal level on challenging international math competitions like the 2025 International Mathematical Olympiad and the 2024 China Mathematical Olympiad. On the notoriously difficult Putnam 2024 exam, it scored 118 out of 120 — far surpassing the top human score of 90. These results place it among the best AI systems for math in the world. What makes this release special is that DeepSeek is sharing it openly. The model’s “weights” — the core of its intelligence — are available for anyone to use and build upon under a permissive license. Tacking the core problem The key problem DeepSeek tackles is that a correct final answer doesn’t always mean the reasoning to get there was correct. Many AI systems are rewarded only for the right answer, which can encourage them to take shortcuts or use flawed logic that isn’t obvious. This is a major issue for tasks like proving theorems or solving open-ended problems, where the step-by-step logic is more important than the final result. How it works To solve this, DeepSeek-Math-V2 uses a clever “prover and reviewer” system that involves two AIs working together. First, the “reviewer” is trained. Before the main model is even trained, a separate AI is taught to act as a proof-checker. It learns to read math solutions, score them on a simple scale, and — crucially — explain which steps are solid and which are flawed, just like a human math grader. Then, the “prover” learns from feedback. The main model, the “prover,” generates solutions. The “reviewer” then scores and critiques them. These scores are used as a reward signal, teaching the prover which reasoning strategies are good and which are bad. The system even includes a “meta-verification” step to check the reviewer itself, ensuring its feedback is accurate and not randomly critical. This keeps the whole learning process honest. This approach encourages the AI to be reflective. A solution that fails but correctly identifies its own mistakes can still get a positive reward, while a wrong answer delivered with overconfidence is penalized. This teaches the AI to be thoughtful rather than to bluff. DeepSeek-Math-V2 is a major upgrade from its predecessor. It’s built on a much larger and more powerful base model and uses this new, sophisticated training method. On technical benchmarks, it outperforms previous open-source models and is competitive with top-tier, specialized systems from other major tech labs. While a dedicated version of Google’s Gemini (Gemini DeepThink) is also at a gold medal level, general-purpose models like GPT-4o and Claude are closer to a silver or bronze level on these advanced proof-based tasks. For developers and researchers, the open availability of DeepSeek-Math-V2 is a significant resource. It provides a practical foundation for building tools that require rigorous, step-by-step reasoning, such as automated theorem provers, homework grading assistants, or specialized AI for scientific research. This model is a clear signal that the next generation of AI will be judged not just on what answers they get right, but on how well they think.(SD-Agencies) | 