Meituan just launched LongCat-2.5, a massive 1.6-trillion-parameter native multimodal model trained entirely on domestic chips. Built for long-horizon agent tasks and visual coding, it activates 48B parameters per token and can generate interactive web pages directly from images. It is now live on its official web chat and available via API.
Meituan just open-sourced LongCat-2.0, a massive 1.6-trillion-parameter MoE model that ditches Nvidia hardware entirely. The pretraining run spanned over 35 trillion tokens on dedicated AI ASIC superclusters without irrecoverable loss spikes, matching frontier models across coding and agent tasks. Why it matters: It proves cutting-edge AI can be trained at scale on alternative chips.
Meituan just open-sourced LongCat-Video-Avatar 1.5, a model that turns a single portrait and audio track into realistic talking video. It handles anime and realistic styles, supports multi-person conversations, and cuts diffusion down to 8 steps for a 15x speedup. The 16GB INT8 weights and code are now available on GitHub and Hugging Face.
Researchers just open-sourced WildActor, a video generation framework that keeps an identity's face, body, and clothing consistent across unconstrained motions using only three reference images. Built on Wan2.2-5B, it supports both real people and cartoon characters while releasing the Actor-18M data construction pipeline.
