Meituan released LongCat-2.5 on November 2, 2026, a new Super Intelligence model from a company better known as China's food delivery giant. This is the full release. A preview had been live on the LongCat API platform since September 25, but today's reports describe a complete native multimodal model, not a preview anymore.
What Meituan is selling is not speed on a single question. LongCat-2.5 is trained for work that runs long: editing a tangled codebase and building a whole project to completion. Meituan says it trained on extended sequences of agent actions, including failed attempts, then used large-scale reinforcement learning to judge the outcome of a whole sequence rather than each reply on its own.
What you can try today
The model is available through the LongCat chat service at longcat.chat and through the LongCat API. Meituan publishes setup guides for Claude Code, Codex, and Hermes, plus OpenCode, Cline, Kilo Code, CodeBuddy, and WorkBuddy. Releases on GitHub and Hugging Face are still promised to follow, so what exists today is a service and an API, not weights you can download and run yourself.
On visual input, Meituan says LongCat-2.5 turns images into tokens at several levels of detail, so text and images move through one shared framework. That is the basis for its claimed support of visual coding, creative design, and 3D tasks. It is a company claim, and we found no independent testing so far to check it against.
Benchmark scores, with a caveat
Meituan's benchmark table gives LongCat-2.5 a score of 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat-2.0, and 51.8 on AutomationBench, up from 15.8. The same table lists 91.9 on BrowseComp, 82.1 on PaperBench, and 78.6 on Vision2Web. But the table also shows the model does not lead everywhere: on Program Bench it scores 25.5, below the top entry of 37 in that comparison. Pandaily, cited by AI Affairs, notes these figures have not been independently verified, so read them as Meituan's claims, not third-party test results.
One figure still in dispute
Several reports describe LongCat-2.5 as a 1.6 trillion parameter model with about 48 billion active per token. That figure needs an asterisk. OrcaRouter reported on October 29, before this full release, that those numbers came from the LongCat-2.0 model card and had been carried across to the preview page. Two reports, different grounds for the same numbers, so we write them as reported figures, not confirmed specs for LongCat-2.5.
Meituan also stresses W4A4 inference optimization, four bits for weights and four bits for activations, on domestic chips. It is a way of using lower numerical precision while the model answers, cutting compute load without losing consistency with training. For readers outside China, the detail matters because it shows the direction: a model built to keep going for a long time, on hardware available at home.
For context, this release lands a week after Mistral opened the weights of Large 4 and Google locked Gemini 4 Carbon to a select group, as in our October 31 update. The pattern cuts three ways: one opens, one closes, one builds for long jobs. And if you want the open-source coding agent that works without being called, compare this with DeepSeek's Super-agent prototype, which is still research code.
Sources: TechNode (Nov 2, 2026), AI Affairs citing Pandaily (Nov 2, 2026), Daily AI Brief (Nov 2, 2026). Figures are snapshots at writing time, afternoon November 2, 2026 WIB, and can change.
Writing about crypto and Super Intelligence: market news, new projects, Super Intelligence models, and early opportunities like airdrops and testnets. Every piece starts from the official announcement, with the costs worked out up front.