Meituan released LongCat-2.5, a 1.6T-parameter multimodal coding model

The Chinese food-delivery firm Meituan has a frontier-scale coding model. The whole training run was done on Chinese AI chips. Meituan released LongCat-2.5 on 2 November 2026. It is a natively multimodal agentic model with 1.6 trillion total parameters.

Meituan says 1.6 trillion parameters are the total count. About 48 billion are active per token. The model has a 1 million token context window. Output can be up to 128,000 tokens per response. All these numbers are Meituan's own. It handles images and text inside one token framework. Meituan says this covers visual coding tasks such as design work and 3D.

The benchmark numbers Meituan published

Meituan reports a Terminal-Bench v2.1 score of 88.4, up from 70.8 for LongCat-2.0. Terminal-Bench measures an agent's success on real command-line tasks. On DeepSWE v1.1, Meituan reports 64.6 for LongCat-2.5 against 9.7 for LongCat-2.0. DeepSWE tests software engineering. On Vision2Web it reports 78.6. On 3DCodeBench, a text-to-3D test, it reports 55.4.

Meituan did not publish head-to-head numbers against Qwen4 Coder, Qwen3.6 or DeepSeek's latest. The scores are all Meituan's own, from its launch material. None has been reproduced by a third party yet.

Chinese chips through the full run

The social card reads: "1.6 trillion total parameters, training done end-to-end on domestic chips." Meituan has not named the chip vendor in English. It did not publish the hardware setup in the launch post. This claim matters for one reason. No Chinese company has stated before that a model this large was trained without any Nvidia GPUs.

Where you can try it

LongCat-2.5 is live on Meituan's own chat service and through its API. Pricing was not published at launch. Rate limits were not disclosed. Meituan says it plans a release on GitHub and Hugging Face. No date was given.

For a team building a coding agent, the number that matters is the DeepSWE jump. Meituan's claim of 9.7 to 64.6 is a 55-point swing over one version. The gap is wide enough that an evaluation harness mismatch between the two runs could partly explain it. The company has not published its evaluation setup. If the number holds against an independent run, LongCat-2.5 joins a small set of models. These models can carry a long software engineering task without a frontier US model behind them.

The thing to watch is the weights release. An API-only model from a food-delivery company is a vendor story. A model on Hugging Face is a tool a reader can run and audit. Until then, Meituan's claims about the training stack and the benchmark scores sit in the "the vendor says" bucket.

Source

Meituan, "Introducing LongCat-2.5". Secondary reporting: TechNode.