Meituan releases LongCat-2.5, a 1.6 trillion parameter model for agentic coding

Meituan released LongCat-2.5 on November 2, Pandaily reports, citing IT Home. The native multimodal model has 1.6 trillion total parameters, activates about 48 billion per token and focuses on long horizon coding and agent tasks.

Meituan says it trained the model with more long horizon trajectory data and reinforcement learning for work such as editing complex codebases and building complete projects. It represents images as discrete tokens that the same backbone processes alongside text, supporting visual coding, design and 3D tasks.

LongCat-2.5 supports Claude Code, Codex, Hermes and other agent tools. It is available through longcat.chat and the LongCat API, while releases on GitHub and Hugging Face are planned.

Meituan reports scores of 88.4 on Terminal Bench v2.1, 51.8 on AutomationBench and 91.9 on BrowseComp. The company also optimized long context training and W4A4 inference, which uses 4 bit weights and activations, for domestic Chinese chips. The model does not lead every test in Meituan's own charts: on Program Bench it scores 25.5, against 37 for the top entry. The benchmark results have not been independently verified.