Meituan open-sources 1.6-trillion-parameter agentic model trained without Nvidia
China's Meituan has released the weights of LongCat-2.5 under the MIT license. It is a model for agents that work through long coding tasks on their own. It has 1.6 trillion parameters in total, about 48 billion of which are active per token, and it was trained without a single Nvidia chip. On the Terminal Bench v2.1 command-line test it scored 88.4 against 89.1 for Claude Opus 5. The figures come from Meituan itself, but teams building agents now have an open option that does not tie them to closed APIs.
AI-processed from The Decoder; edited by Hamidun News
The company most people know for food delivery has open-sourced one of the largest agentic models available. Meituan published the weights of LongCat-2.5: 1.6 trillion parameters, an MIT license, and training done entirely on Chinese hardware. For anyone building coding agents, that sends two signals at once: open models have again closed in on the level of closed ones, and Nvidia is no longer a prerequisite.
What the model is and how it works
LongCat-2.5 uses a mixture-of-experts architecture. Think of a large medical consultation where, for each question, only a few relevant specialists are called in rather than every doctor. The model works the same way: of its 1.6 trillion parameters, about 48 billion are active for each token. That makes it noticeably cheaper to run than a dense model of the same size.
The API offers a context window of up to a million tokens, and a response can run to 128 thousand tokens. The model understands text, images and video and replies in text, and reasoning can be switched on and off. Compared with LongCat-2.0, it adds image understanding: images are split into discrete tokens at native resolution and processed together with text, without a separate external encoder. Meituan expects this to help with visual coding, design and 3D tasks.
For reinforcement learning, the team used real tasks from GitHub pull requests along with reproduced agent errors. To keep the model from cheating, they disabled internet access in the environment, removed Git history and caches, and asked another language model to check solutions for tricks. The skills of several specialized teacher models, covering agentic work, reasoning and dialogue, were distilled into a single one.
Hardware: 194 Huawei chips
The most unusual part is the infrastructure. Meituan quantizes the weights and activations of the experts to 4 bits, which Chinese accelerators can process natively. In internal load tests, throughput rose by 43.3% for prompt processing and by 42.77% for response generation compared with the 8-bit version. According to the company, accuracy stayed on par with the unquantized model. This version sits in a separate repository.
The self-hosting guide describes a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series: two handle images, 64 process prompts, and 128 generate responses. There are no instructions for Nvidia. The predecessor, LongCat-2.0, was trained in June on a cluster of more than 50,000 Chinese chips, and before its release it ran anonymously on OpenRouter under the name Owl Alpha.
What the benchmarks show and where the catch is
On Terminal Bench v2.1, which tests how well an agent handles command-line tasks, LongCat-2.5 scored 88.4. In Meituan's table, Claude Opus 5 gets 89.1 and GPT-5.6 Sol gets 88.8, while Deepseek-V4.1-Flash leads with 90.6. On DeepSWE, where models fix bugs in real repositories, the result rose from 9.7 to 64.6 points, but that is still below everything else in the table except Claude Opus 4.8; the leader is Deepseek-V4.1-Flash with 74.2. On the harder Terminal Bench v4.0 it scores 38.5: slightly above GPT-5.6 Sol (37.3) but far from Opus 5 (53.9). On the BrowseComp web-search test the model scored 91.9, the best result in the table, but Meituan ran that test in its own environment.
Honest caveats: all the figures come from the company itself, and the table leaves out the newer Claude Opus 5.5 and GPT-6.1 Sol. In terms of capability, this is more "what top models could do a few months ago."
What this means in practice
For developers, the terms matter most. The MIT license allows commercial use, modification and redistribution. The API is compatible with OpenAI and Anthropic formats and connects to Claude Code, Codex and Hermes, so you can try it with almost no changes to your tooling. A promotional price is currently in effect: about €0.27 per million input tokens and €1.07 per million output tokens. There is also a web chat at longcat.ai.
But self-hosting a model of this size means a cluster of well over a hundred accelerators, and on Ascend rather than the familiar Nvidia. For most teams, the API is the more realistic route. For the industry, something else matters more: open Chinese models such as Kimi K3.1 and GLM-5.3 have nearly caught up with the top on popular tests, and when they are optimized for local chips from the start, that makes such hardware more attractive outside China too.
- Nov 18, 2026Together AI replaces megakernels with AI agents and speeds up Llama 1B by 28% on B200
- Nov 18, 2026A $3,000 Mining Card Turned Into an A100 Equivalent: From 8 to 80 GB of Memory
- Nov 12, 2026Jensen Huang steps down as Nvidia CEO: company loses $1.4 trillion in two weeks
- Nov 12, 2026Two engineers integrated SGLang with IBM Spyre in a week and boosted throughput 3.6-fold
Key Questions Answered
What the model is and how it works?
LongCat-2.5 uses a mixture-of-experts architecture. Think of a large medical consultation where, for each question, only a few relevant specialists are called in rather than every doctor. The model works the same way: of its 1.6 trillion parameters, about 48 billion are active for each token. That makes it noticeably cheaper to run than a dense model of the same size. The API offers a context window of up to a million tokens, and a response can run to 128 thousand tokens.
What are the key details on Hardware: 194 Huawei chips?
The most unusual part is the infrastructure. Meituan quantizes the weights and activations of the experts to 4 bits, which Chinese accelerators can process natively. In internal load tests, throughput rose by 43.3% for prompt processing and by 42.77% for response generation compared with the 8-bit version. According to the company, accuracy stayed on par with the unquantized model. This version sits in a separate repository.
What the benchmarks show and where the catch is?
On Terminal Bench v2.1, which tests how well an agent handles command-line tasks, LongCat-2.5 scored 88.4. In Meituan's table, Claude Opus 5 gets 89.1 and GPT-5.6 Sol gets 88.8, while Deepseek-V4.1-Flash leads with 90.6. On DeepSWE, where models fix bugs in real repositories, the result rose from 9.7 to 64.6 points, but that is still below everything else in the table except Claude Opus 4.8; the leader is Deepseek-V4.1-Flash with 74.2.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.