Meituan releases LongCat-2.5: a multimodal coding model built for long, multi-step engineering work

Meituan released LongCat-2.5 on November 2, 2026, a native multimodal model built for coding and agent work that stretches across many steps. The model is aimed at sustained execution by AI agents: editing a complex codebase, or carrying a software project through to completion, rather than answering a single prompt. It is available through Meituan’s online chat service and API platform, with releases on GitHub and Hugging Face planned to follow.

The release follows a preview. LongCat-2.5-Preview went live on the LongCat API platform on September 25, 2026, and the November 2 release is the fuller launch: Meituan now attaches a benchmark table to the model, describes expanded training on long task sequences, and says it has optimised low-precision inference for domestic chips. AI Affairs reports the release as moving from that September preview to the full model, citing Pandaily.

Trained on long sequences of agent work#

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, the model receives feedback on the outcome of a whole sequence of actions, rather than treating each response as a separate exercise. The company also says it turns failures encountered by real agents into reproducible training tasks, so the training process includes examples of work that went wrong and can be checked.

For visual input, LongCat-2.5 converts images into tokens at several levels of detail, allowing its main processing network to handle visual material and text in token form. Meituan connects that design to visual coding, creative design and 3D tasks, and describes synthesised data, constructed environments and mixed training across different agent harnesses — the software that lets a model take actions during a task.

The numbers come with two caveats#

Meituan’s benchmark table scores LongCat-2.5 at 88.4 on Terminal Bench v2.1, compared with 70.8 for LongCat-2.0, and 51.8 on AutomationBench, compared with 15.8 for its predecessor. The same table lists 91.9 on BrowseComp, 82.1 on PaperBench, 55.4 on 3DCodeBench and 78.6 on Vision2Web. It does not lead everywhere: on Program Bench it scores 25.5, against 37 for the top entry in the comparison. Pandaily reported that the benchmark figures had not been independently verified, and the scores are Meituan’s own measurements, not a record of how often a coding task succeeds in ordinary use.

The reported size carries a separate caveat. Pandaily and TechNode give LongCat-2.5 a total of 1.6 trillion parameters, with about 48 billion activated for each token. OrcaRouter reported on October 29, before the full release, that those figures 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, so the parameter count should be read as reported, not settled.

Four-bit inference on 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 — a way of using less numerical precision during inference, when a trained model generates an answer. The company says it worked on keeping those calculations consistent with training, and on handling long sequences of agent actions during training, in support of 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. It matters to the release in the same way the training does: a model meant to keep working across extended tasks needs the hardware running each successive step to keep up.

What to watch#

The practical test is whether developers can put the model to work through the tools they already use. The LongCat change log says LongCat-2.5 works with Claude Code, Codex and Hermes, and Meituan’s platform documentation includes setup guides for OpenCode, Cline, Kilo Code, CodeBuddy and WorkBuddy. The planned GitHub and Hugging Face releases would settle how the model is distributed beyond the API; until they appear, availability means the chat service and the API platform.

The other open question is independent verification. The benchmark table is a company claim, and the parameter figures are disputed. Both can be checked once outside groups run the model on the same tests — and that check, more than the launch-day table, will decide whether LongCat-2.5 is a step forward for long-horizon coding or a preview with a new score sheet.