Meituan fully released LongCat-2.5 on November 2, 2026 — a native multimodal model built for long-horizon coding and agent work, with a reported 1.6 trillion total parameters and about 48 billion active per token. Why it matters: it is another sign that China’s biggest consumer platforms are now shipping frontier-scale models tuned for autonomous software work, not just chat.
November 2, 2026 — Beijing. Meituan’s LongCat team has moved LongCat-2.5 from the API preview it opened on September 25 to a full release, available through the longcat.chat service and the LongCat API platform, with GitHub and Hugging Face releases planned to follow.
What happened
LongCat-2.5 processes text and images inside a single, unified token framework, so the same model that reads a codebase can also work from screenshots and visual inputs. Meituan positions it for visual coding, creative design and 3D tasks, and for agent jobs that run across many steps — editing complex codebases and completing whole software projects rather than answering one prompt at a time. Reporting on the launch says the model works with agent tools including Claude Code, Codex and Hermes.
The details
Meituan says it expanded training on long task sequences and applied large-scale reinforcement learning aimed at sustained execution. One notable training trick: failures that real agents hit in the wild are converted into reproducible training tasks, so the model practices on the exact situations where agents tend to break down.
On Meituan’s own benchmark table, LongCat-2.5 scores 88.4 on Terminal Bench v2.1 (up from 70.8 for LongCat-2.0), 51.8 on AutomationBench and 91.9 on BrowseComp. Those numbers are company-reported and have not been independently verified — and Meituan’s own charts show the model does not lead everywhere: on Program Bench it scores 25.5 against 37 for the top entry. The headline size figures also follow the LongCat-2.0 lineage, so treat the 1.6-trillion-parameter count as Meituan’s reported spec for the family rather than a newly audited number.
Why it matters
The interesting shift here is what the model is optimized for. A year ago, Chinese model launches competed on chat quality and price. LongCat-2.5 is pitched squarely at long-running agent work — the same territory Anthropic’s Claude Code and OpenAI’s Codex are fighting over — and it comes from a food-delivery and local-services giant, not a pure AI lab. If Meituan’s reported terminal and automation scores hold up under independent testing, Western coding-agent vendors have one more serious, low-cost rival to reckon with. The planned open releases on GitHub and Hugging Face will be the real test: once the weights and reproducible evals are public, the claims can be checked instead of taken on trust.
FAQ
What is LongCat-2.5?
It is Meituan’s latest large AI model: a native multimodal system that takes text and image inputs in one framework, reported at 1.6 trillion total parameters with roughly 48 billion active per token, and tuned for long-horizon coding and agent tasks.
Where can you use LongCat-2.5?
It is available now through Meituan’s longcat.chat service and the LongCat API platform. Releases on GitHub and Hugging Face are planned but had not landed at launch.
Are LongCat-2.5’s benchmark scores verified?
No. The headline scores — 88.4 on Terminal Bench v2.1, 51.8 on AutomationBench, 91.9 on BrowseComp — come from Meituan’s own benchmark table and have not been independently verified.
How is this different from the September preview?
LongCat-2.5 first appeared as an API preview on September 25, 2026. The November 2 release is the full launch, with expanded long-sequence training and reinforcement learning for complex, multi-step coding work.
Sources: TechNode, Longbridge, Pandaily via Daily AI Brief and AI Affairs.