要点速览
- DeepMind's AlphaProof Nexus solves multiple long-standing mathematical problems using Gemini
- Hugging Face releases multiple high-impact open-source AI models and datasets
- Yann LeCun shares JEPA-WM v2 research with formal publication and reproducibility certification
- NVIDIA showcases local AI agent deployment capabilities on GB10 hardware
一、AI Mathematical Research Breakthroughs
- DeepMind's Demis Hassabis announced AlphaProof Nexus, a Gemini-based intelligent agent framework for formal mathematical proof search. The system autonomously solved 9 open Erdős problems (including two unsolved for 56 years), 44 OEIS problems, one 15-year unsolved algebraic geometry problem, and one 7-year unsolved minimax optimization problem. The team is collaborating with multi-disciplinary mathematicians and released the associated research paper. — via 1
二、Open-Source AI Model and Dataset Updates from Hugging Face
- Hugging Face announced the MiniCPM5-1B fully open-source 1B parameter model, which ranks first among all sub-2B parameter models on Artificial Analysis with a score of 17.9, outperforming Qwen3.5-2B (16.3) and other comparable models. It supports INT4 quantization to 0.5GB for deployment on mobile, browser and edge devices, and was trained using the AI-written ForgeTrain framework which is 10% faster than NVIDIA Megatron without human programmer involvement. — via 1
- The platform released MiMo V2.5-Coder model, which outperforms Qwen 3.6 and DeepSeek 4-Flash in author experiments and is optimized for users with 128GB of memory. — via 1
- Launched NayanaOCR, the largest open-source synthetic multilingual multimodal document corpus with over 1 million document images covering 22 languages. — via 1
- Announced that llama.cpp with MTP support delivers a 78% speedup for local model inference, reaching 45 tok/s on A10G GPUs for Qwen3.6-27B dense generation tasks, making it suitable for daily use. — via 1
三、Academic and Technical Research Releases
- Yann LeCun announced the release of JEPA-WM v2, which has been accepted by TMLR and awarded reproducibility certification. The updated version adds data scaling experiments, Lipschitz analysis for multi-step unfolding training, and expanded discussions, with authors including Jimmy Yang, Jean Ponce, Adrien Bardes and LeCun himself. — via 1
四、Industry Deployment and Trend Updates
- NVIDIA AI demonstrated that 2x DGX Sparks with MiniMax M2.7 NVFP4 can run 16 local AI agents simultaneously using only 2 GB10 chips without cloud API reliance. — via 1
- Ethan Mollick noted that as more users become familiar with AI-generated content characteristics, it will become increasingly clear that large portions of current platform content, blog posts and papers are AI-generated. — via 1
