Meituan released LongCat-2.5 on 2 November 2026, moving from a September API preview to a native multimodal model intended for coding tasks that run across many actions, Pandaily reported. The release puts Meituan’s claims about extended engineering work alongside benchmark scores and an optimisation for running the model on domestic chips. LongCat-2.5 is available through the LongCat API and at longcat.chat.
Key points
- Meituan says it trained the model on extended sequences of agent actions, including failed attempts turned into training tasks.
- Its benchmark table gives LongCat-2.5 scores of 88.4 on Terminal Bench v2.1 and 51.8 on AutomationBench.
- Meituan says it optimised 4-bit weights and activations for inference on domestic chips.
- The reported 1.6 trillion total parameters and approximately 48 billion active per token are disputed by a report about the earlier preview.
LongCat-2.5 trains on extended coding runs
Meituan says it expanded training data drawn from long sequences of agent actions and used large-scale reinforcement learning for work inside complex codebases and for building complete projects. In that training method, the system receives feedback on the outcome of a sequence of actions, rather than treating each response as a separate exercise. The company also says it turns failed runs by real agents into tasks that can be repeated and checked, giving the training process examples of work that went wrong.
Editing an existing piece of software could benefit from a system trained to carry work through a complex codebase rather than stop at a suggested snippet. Building a complete project would call for the same continuity over a longer task. Those are the kinds of engineering work Meituan says it targeted in training LongCat-2.5.
For visual input, Meituan says LongCat-2.5 converts images into tokens at several levels of detail, allowing the model’s main processing network to handle visual material and text in token form. The company connects that design to visual coding, creative design and 3D tasks. Its account of the training also describes synthesised data, constructed environments and mixed training across different agent harnesses — the software that lets a model take actions during a task.
Pandaily gives LongCat-2.5 a total of 1.6 trillion parameters, with about 48 billion used for each token it processes. OrcaRouter reported on 29 October, before the full release, that the figures attached to LongCat-2.5 came from the LongCat-2.0 model card and had been carried across to the preview. The two accounts give different grounds for attaching the same figures to the new model.
Meituan optimises W4A4 for domestic chips
Meituan says it optimised long-context training and W4A4 inference on domestic chips. W4A4 uses four bits to represent each model weight and four bits for each activation, the intermediate values produced while the model processes a request. It is a way of using less numerical precision during inference — the stage when a trained model generates an answer — so keeping the calculations consistent with training matters. Meituan says it worked on that numerical consistency and on the handling of long sequences of agent actions during training.
The domestic-chip work matters to the release because the model is intended to keep working across extended tasks, while the hardware must run each successive step. Meituan describes its optimisations as supporting large-scale, long-horizon reinforcement learning. That is the company’s account of its engineering work, rather than a reported measurement of inference speed or performance under production load.
Meituan reports 88.4 on Terminal Bench
Meituan’s benchmark table scores LongCat-2.5 at 88.4 on Terminal Bench v2.1, compared with 70.8 for LongCat-2.0. On AutomationBench, the table gives the new model 51.8 and its predecessor 15.8. The results are scores on those named tests, reported by the company, rather than measurements of how often a coding task succeeds in ordinary use.
The same table lists 91.9 on BrowseComp, 82.1 on PaperBench, 55.4 on 3DCodeBench and 78.6 on Vision2Web. It gives LongCat-2.5 a score of 25.5 on Program Bench, against 37 for the top entry in that comparison. Pandaily reported that the benchmark figures had not been independently verified.
LongCat API follows September preview
LongCat-2.5-Preview went live on the LongCat API platform on 25 September. In its account of the preview, OrcaRouter described a difference between the platform’s image-capable change log and an API reference that listed text as the input modality. That report preceded the 2 November release and concerned the preview documentation. For developers considering visual tasks, the documented input format is a separate question from the model’s reported visual training.
The LongCat change log says LongCat-2.5 works with Claude Code, Codex and Hermes, giving developers existing harnesses through which to use the released model. Meituan’s platform documentation also includes setup guides for OpenCode, Cline, Kilo Code, CodeBuddy and WorkBuddy.
