Open models and agent systems lead AI updates

Open-source model activity accelerated as Moonshot AI open-sourced Kimi K3, described as the world’s first 3-trillion parameter level open-source model with actual 2.8T parameters and 1 million token ultra-long context. Ant Group released LLaDA2.2, an Agentic diffusion model that supports four atomic editing operations, scores 53.83 on average across 7 Agent benchmarks and delivers 1.64x the throughput of comparable autoregressive models.

AI infrastructure remained a major funding and deployment theme. SSI formed a strategic partnership with NVIDIA, which will invest approximately $5 billion and provide the Vera Rubin computing platform, enabling SSI’s computing power to increase 10-fold within 12 months. Bitcoin miners TeraWulf and Hut 8 are shifting toward AI data centers, with TeraWulf signing a $19 billion 20-year lease with Anthropic and Hut 8 securing a $9.8 billion 15-year contract.

Agent and robotics work expanded across research and products. Peking University, Tsinghua University, Beihang University, Shanghai Jiao Tong University and University of Science and Technology of China launched TriWorldBench Challenge with 500 tri-view synchronized episodes covering 50 robotic manipulation tasks. Xu Xin’s venture framework highlighted likely LLM consolidation to 3-5 players, embodied AI penetration at 2-3%, and the need to back winners for 8-10 years.