Meituan releases LongCat-2.5, a 1.6-trillion-parameter agent model trained on Chinese hardware

The LongCat team at Chinese food delivery company Meituan has released the weights for LongCat-2.5, one week after the model's official unveiling. It is built for agents that work through long, multi-step tasks autonomously, particularly coding. According to Meituan, it comes close to top Western models on several benchmarks. The article says the model was trained entirely on Chinese hardware.

LongCat-2.5 is a mixture-of-experts model, so only a subset of its specialized subnetworks runs for each token. Of 1.6 trillion total parameters, roughly 48 billion are active per token. Through the API it handles up to one million tokens at once and generates responses up to 128,000 tokens long. It understands text, images and video, responds in text, and lets users toggle reasoning on and off.

The model builds on LongCat-2.0 and adds image understanding. Rather than bolting on a separate image encoder, it converts images into tokens that the model processes alongside text. The approach draws on the research model LongCat-Next and aims to improve visual coding, design and 3D tasks. For reinforcement learning, the team used real software tasks from GitHub pull requests plus recreated agent failures. To stop the model from cheating, the developers blocked internet access in the training environments, removed Git history and cached files, and had a language model check solutions for tricks. The team also distilled the capabilities of several specialized teacher models for agent work, reasoning and dialog into the model.

Meituan quantizes the experts' weights and activations to 4-bit precision, which Chinese chips can process natively. In Meituan's own load tests, throughput rises 43.3 percent for prompt processing and 42.77 percent for response generation compared with an 8-bit version, and Meituan says accuracy stays on par with the uncompressed model. This version sits in a separate repository. Its self-hosting guide describes a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series: two handle image processing, 64 handle prompt ingestion and 128 handle response generation. There is no guidance for Nvidia GPUs.

On benchmarks, as listed by Meituan: on Terminal Bench v2.1, which measures command-line agent tasks, LongCat-2.5 scores 88.4, against 89.1 for Claude Opus 5, 88.8 for GPT-5.6 Sol and 90.6 for Deepseek-V4.1-Flash, which tops the table. On DeepSWE, which tests bug fixing in real repositories, it jumps from 9.7 to 64.6, behind every other model in the table except Claude Opus 4.8; Deepseek-V4.1-Flash leads at 74.2 and GPT-5.6 Sol scores 72.7. The newer Claude Opus 5.5 and GPT-6.1 Sol are not in the table. On the harder Terminal Bench v4.0, LongCat-2.5 scores 38.5, slightly above GPT-5.6 Sol's 37.3 but well below Opus 5's 53.9. It posts 91.9 on the web search benchmark BrowseComp, the highest in the table, though Meituan ran that test in its own environment.

The weights are under the MIT license, which allows commercial use, modification and redistribution. A chat interface is at longcat.ai. The API supports both the OpenAI and Anthropic formats and plugs into agent harnesses such as Claude Code, Codex and Hermes. At launch Meituan charges a limited-time promotional price equivalent to roughly €0.27 per million input tokens and €1.07 per million output tokens.

For context, Meituan says it trained the predecessor LongCat-2.0 in late June entirely on a cluster of more than 50,000 Chinese AI chips; before that, the model ran anonymously as "Owl Alpha" on OpenRouter. The article notes that Chinese open-weight models like Kimi K3.1 and GLM-5.3 have largely caught up with top US models on common benchmarks, and that open-weight models optimized for Chinese chips from the start make that hardware more attractive outside China too. Meituan is also working on image generators; it previously released the open-source text-to-image model LongCat-Image, which uses just 6 billion parameters.

Key facts

  • LongCat-2.5 is a mixture-of-experts model with 1.6 trillion total parameters, roughly 48 billion active per token, and a context of up to one million tokens through the API.
  • Weights are MIT-licensed; the article says the model was trained entirely on Chinese hardware, and the self-hosting guide covers 194 Huawei Ascend Atlas A2 accelerators with no Nvidia guidance.
  • Meituan's table has it at 88.4 on Terminal Bench v2.1 (Claude Opus 5 89.1, GPT-5.6 Sol 88.8, Deepseek-V4.1-Flash 90.6) and 64.6 on DeepSWE, behind all other listed models except Claude Opus 4.8.
  • It scores 91.9 on BrowseComp, the highest in the table, but Meituan ran that test in its own environment; on Terminal Bench v4.0 it gets 38.5 against Opus 5's 53.9.
  • A 4-bit quantized version raises throughput 43.3 percent for prompt processing and 42.77 percent for response generation versus 8-bit in Meituan's tests; launch API pricing is roughly €0.27 input and €1.07 output per million tokens, for a limited time.

Why it matters

A 1.6-trillion-parameter open-weight agent model, released under a permissive license and described as trained entirely on Chinese hardware, shows a large model can be built and deployed without Nvidia GPUs. The self-hosting guide targets Huawei Ascend accelerators, and the 4-bit quantization is chosen because Chinese chips process it natively. The article argues that open-weight models optimized for Chinese chips from the start make that hardware more attractive outside China too. It also places the release in a wider trend: Chinese open-weight models like Kimi K3.1 and GLM-5.3 have largely caught up with top US models on common benchmarks.

Who it affects

Teams building coding and long-horizon agents who want an open-weight model they can self-host or call through an API. Operators with Huawei Ascend hardware get a documented deployment layout. Developers already using Claude Code, Codex or Hermes can point those harnesses at the API, which supports the OpenAI and Anthropic formats. Companies that want commercial use and redistribution rights get them under the MIT license.

How to use it

The weights are available under the MIT license, which allows commercial use, modification and redistribution. You can try the model in a chat interface at longcat.ai. The API supports the OpenAI and Anthropic formats and works with agent harnesses such as Claude Code, Codex and Hermes. At launch Meituan charges a limited-time promotional price equivalent to roughly €0.27 per million input tokens and €1.07 per million output tokens. For self-hosting, a 4-bit quantized version is in a separate repository, and its guide describes 194 Huawei Ascend Atlas A2 accelerators: 2 for image processing, 64 for prompt ingestion and 128 for response generation.

How solid is it

The benchmark figures are as listed by Meituan; the article reports no independent verification. BrowseComp, where the model posts the table's highest score of 91.9, was run in Meituan's own environment. The 4-bit throughput gains come from Meituan's own load tests, and Meituan says accuracy matches the uncompressed model. The source does not say what the 9.7 DeepSWE figure refers to. The claim of training entirely on Chinese hardware is stated by the article; the 50,000-plus Chinese chips figure is Meituan's statement about LongCat-2.0, and the size of the training cluster for LongCat-2.5 itself is not given.

Risks and caveats

Results are mixed against the leaders. On DeepSWE the model sits behind every other listed model except Claude Opus 4.8, and the newer Claude Opus 5.5 and GPT-6.1 Sol are not in the table. On Terminal Bench v4.0 it scores 38.5 against Opus 5's 53.9. On Terminal Bench v2.1 and DeepSWE a Deepseek model tops the table. The self-hosting guide offers no guidance for Nvidia GPUs, so anyone off Ascend hardware has less to go on. The promotional API price is limited-time, and the article gives no end date or post-promotion price.

“There's no guidance for Nvidia GPUs.”

— The Decoder, on the LongCat-2.5 self-hosting guide