第1个获得数学奥赛金牌的开源模型！DeepSeek新模型获网友盛赞：公开技术文件，了不起！

Core Insights - DeepSeek has launched its latest mathematical reasoning model, DeepSeekMath-V2, which has achieved gold medal status in the simulated 2025 International Mathematical Olympiad (IMO), marking a significant breakthrough in open-source AI's complex reasoning capabilities [1][2] - This achievement positions DeepSeekMath-V2 as the first open-source model to win a gold medal at the IMO level, drawing attention from the AI research and developer community [2] - Unlike closed-source models from Google and OpenAI, DeepSeekMath-V2's model weights are publicly available under the Apache 2.0 license, allowing for unrestricted access and exploration by users [3][5] Performance Highlights - DeepSeekMath-V2 solved 5 out of 6 problems in the IMO 2025 simulation, achieving gold medal status, which is a notable accomplishment given that only 72 out of 630 human participants received gold medals [4] - The model also demonstrated top-tier performance in other prestigious competitions, including achieving gold medal status in the Chinese Mathematical Olympiad (CMO) and scoring 118 out of 120 in the Putnam Mathematics Competition [4] Open-Source Advantage - The core appeal of DeepSeekMath-V2 lies in its complete openness, allowing users to freely download, fine-tune, and optimize the model without restrictions [5] - The release has been praised as a significant milestone for the open-source community, emphasizing the potential for open-source models to challenge the commercial strongholds of closed-source products [3][5] Innovative Training Framework - DeepSeekMath-V2 employs an innovative self-verification training framework, which includes a specialized verifier that assesses the quality of the proof process rather than just the final answer [10][11] - This mechanism encourages the model to identify and rectify issues in its reasoning chain before finalizing answers, enhancing the rigor of its mathematical reasoning [12] Dynamic Evolution Strategy - To prevent overfitting to its own verification mechanism, DeepSeek has implemented a dynamic evolution strategy that increases computational demands and automatically labels difficult proofs, ensuring the verifier and generator evolve in tandem [13] - This approach allows for the continuous optimization of performance and the creation of new training data, validating the feasibility of self-driven learning systems in tackling complex mathematical reasoning tasks [13]