On November 9, Meituan announced the open-source release of LongCat-2.5 under the MIT license, aimed at long-horizon agent tasks as a native multimodal model. It totals 1.6 trillion parameters with about 48 billion activated per token and supports a one-million-token context.
The model is co-optimized for domestic compute with limited memory and bandwidth: when a single card has under 80 GB, sharding spreads the load across super-node cards, and long-sequence supervised fine-tuning throughput rose 86% over the DDP baseline. It marks a substantive "home chip, home model" breakthrough.
The release supports commercial inference deployment and secondary training without extra authorization, and its end-to-end delivery improves in complex codebases, full-project builds, visual coding and professional office scenarios.
Source: Meituan / WeChat
