Meituan releases LongCat-2.5, agent model trained on Chinese hardware

Meituan has published the weights of LongCat-2.5, a 1.6 trillion parameter model designed for agents. The company states it was trained on Chinese hardware, without mentioning Nvidia, targeting advanced code and automation uses.

Meituan has released the weights of LongCat-2.5, a 1.6 trillion parameter AI model designed for agents that chain complex tasks. According to the Chinese company, training was conducted on Chinese hardware, with no mention of Nvidia GPUs. This topic is significant as it combines performance, hardware sovereignty, and commercial openness, with a model already close to the best Western systems on several benchmarks.

A model tailored for long tasks, not for flashy demos

LongCat-2.5 is a mixture-of-experts type model, meaning it activates only part of its subnetworks for each token processed. Out of 1.6 trillion parameters total, about 48 billion are active per token. Meituan provides a context window of up to one million tokens via the API, with responses up to 128,000 tokens. The model understands text, images, and video, then responds in text, with an option to enable or disable reasoning.

This orientation mainly targets software agents, i.e., systems capable of executing sequences of actions without human intervention at each step. Meituan says it trained LongCat-2.5 on real tasks from GitHub pull requests, while blocking usual shortcuts in the test environment: internet access cut, Git history deleted, cached files removed. The language model responsible for verifying responses also had to detect cheating tricks. It’s not glamorous, but often this is where the difference between a demo and a useful tool lies.

Des résultats proches des leaders américains sur plusieurs benchmarks

On Terminal Bench v2.1, LongCat-2.5 scores 88.4 points, behind Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8, according to figures published by Meituan. Deepseek-V4.1-Flash tops this chart with 90.6 points. On Terminal Bench v4.0, the more difficult version, the model reaches 38.5 points, ahead of GPT-5.6 Sol at 37.3, but far from Opus 5 at 53.9. Meituan also reports 91.9 points on BrowseComp, its best score in published tests, but this result was obtained in its own environment [to be verified].

The clearest progress appears on DeepSWE, a test dedicated to bug fixing in real code repositories. LongCat-2.5 jumps from 9.7 to 64.6 points, according to Meituan’s internal comparison. The leap is clear. Benchmarks never tell the whole story, but here they give one thing clearly: the model has reached a new level on software engineering tasks.

Le pari du matériel chinois change la donne industrielle

Meituan also highlights a reference deployment on 194 Huawei Ascend Atlas A2 accelerators. Two are used for image processing, 64 for prompt ingestion, and 128 for response generation. No Nvidia recommendations appear in the self-hosting guide. In a context where supply chains remain sensitive, this detail is anything but decorative.

The company says it quantized the weights and expert activations in 4 bits, a more compact format that improves throughput on these chips. In load tests, throughput increases by 43.3% for prompt processing and 42.77% for response generation compared to an 8-bit version, while maintaining comparable accuracy according to Meituan. The company also published the weights under an MIT license, allowing commercial use, modification, and redistribution.

Une diffusion large, et un signal pour le marché des modèles ouverts

LongCat-2.5 is accessible via a chat interface on longcat.ai and via an API compatible with OpenAI and Anthropic formats. The model also integrates with agent tools like Claude Code, Codex, and Hermes. At launch, Meituan charges a promotional price of about 0.27 euro per million input tokens and 1.07 euro per million output tokens. The company had already presented LongCat-2.0, trained at the end of June on more than 50,000 Chinese AI chips, after a testing period under the name “Owl Alpha” on OpenRouter.

The scope of the launch thus goes beyond the technical sheet alone. By showing that an advanced agent model can be trained and deployed on Chinese hardware, Meituan sends a message to the market: open weight is no longer just a matter of model size, but also of available infrastructure. For developers and companies, this opens an additional option. For chip suppliers, too.

Key points

  • 1.6 trillion parameters for LongCat-2.5.
  • 48 billion active parameters per token.
  • 194 Huawei Ascend chips in the typical deployment.
  • 88.4 points on Terminal Bench v2.1.
  • MIT license for commercial use.
  • Meituan released LongCat-2.5 on November 15, 2026.

By the numbers

  • 1.6 trillion parameters — total size of the LongCat-2.5 model, according to Meituan.
  • 48 billion active parameters — portion activated per token in this mixture-of-experts architecture.
  • 194 Huawei Ascend chips — reference cluster configuration described by Meituan.
  • 88.4 points — LongCat-2.5’s score on Terminal Bench v2.1, according to Meituan.
  • 0.27 euro / 1.07 euro per million tokens — promotional prices announced at launch.

To read