腾讯研究院AI速递 20260122

Group 1 - DeepSeek's Model 1 has been discovered in the FlashMLA codebase, potentially indicating an upcoming release, featuring a 512-dimensional architecture and support for NVIDIA's Blackwell architecture [1] - Liquid AI has launched the open-source inference model LFM2.5-1.2B-Thinking, which operates on a liquid neural network architecture and requires only 900MB of memory on mobile devices, achieving a score of 88 on MATH-500 [2] - The xAI engineer revealed that AI is being tested as a "colleague" in the MacroHard project, achieving human speeds eight times faster, and the company is considering utilizing idle computing power from approximately 4 million Tesla vehicles in North America [3] Group 2 - Research indicates that models like DeepSeek-R1 can spontaneously form multi-role debate mechanisms, significantly improving accuracy through internal social dialogue [4][5] - Medical SAM3, a new model developed by the University of Central Florida, allows for expert-level segmentation in medical imaging using only text prompts, achieving an average accuracy increase from 11.9% to 73.9% across 33 datasets [6] - Anthropic's CEO predicts that AI will fully take over software engineering roles within 6-12 months, with a significant portion of entry-level jobs expected to disappear in the next 1-5 years [7] Group 3 - The Sequoia xbench team reported that top agents can handle over 60% of 104 daily tasks, indicating that foundational agent capabilities have become commoditized [8] - OpenAI's CFO discussed the maturation of multi-agent systems by 2026, emphasizing that AI bubbles should be measured by API call volumes rather than stock prices, with productivity increases of 27-33% for cutting-edge companies [9]