Models

Meituan open-sources 1.6-trillion-parameter agent model trained without Nvidia

Meituan open-sources LongCat-2.5, a 1.6-trillion-parameter agent model trained entirely on Chinese chips, scoring 88.4 on Terminal Bench v2.1 under an MIT license.

Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU
Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPUColossusCloud / pixabay
By Sophie Lindqvist5 min read

Why it matters

  • Meituan released LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model (48B active per token), under the MIT license.
  • The model was trained entirely on Chinese hardware; predecessor LongCat-2.0 used a cluster of more than 50,000 Chinese AI chips.
  • LongCat-2.5 scores 88.4 on Terminal Bench v2.1, near Claude Opus 5's 89.1.
  • The self-hosting guide specifies 194 Huawei Ascend Atlas A2 accelerators, with no Nvidia guidance.
  • Promotional API pricing is roughly €0.27 per million input tokens and €1.07 per million output tokens.

Meituan has released the weights of LongCat-2.5, a 1.6-trillion-parameter model built for autonomous AI agents, and the entire system was trained on Chinese hardware without a single Nvidia GPU. One week after the model's official unveiling, the food-delivery giant's LongCat team published the weights under an MIT license, alongside benchmark results that put the model close to top US systems on several coding and agent tests.

The release matters beyond one company's leaderboard position. It's concrete evidence that a Chinese firm can train a frontier-class agent model end to end on domestic chips and then give it away with commercial rights — a combination that pressures both US model providers and Nvidia's export position at the same time.

What is LongCat-2.5?

LongCat-2.5 is a mixture-of-experts model. Only a subset of its specialized subnetworks fires for each token it processes: out of 1.6 trillion total parameters, roughly 48 billion are active per token. That's the same architectural trick that lets very large models run at manageable cost.

Through the API, the model handles up to one million tokens of context 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 team built it for agents that work through long, multi-step tasks autonomously, with coding as the headline use case.

How does it handle images?

LongCat-2.5 builds on its predecessor LongCat-2.0 and adds image understanding. Many multimodal models bolt a separate image encoder onto the model after the fact. LongCat-2.5 instead converts images into tokens that the model processes alongside text, which is standard practice for today's large language models.

The approach draws on the research model LongCat-Next and aims to improve visual coding, design, and 3D tasks. In short, the model sees images the same way it sees words.

How did Meituan train it?

For reinforcement learning, the team used real software tasks from GitHub pull requests along with recreated agent failures. To make sure the model actually solved the tasks rather than gaming them, the developers blocked internet access in the training environments, removed Git history and cached files, and had a language model check its solutions for tricks.

The team also distilled the capabilities of several specialized teacher models for agent work, reasoning, and dialog into LongCat-2.5. So the final model is partly a compression of expert systems, partly a product of trial and error on real code.

What about the hardware?

This is the part of the story with the widest stakes. 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.

For deployment, Meituan quantizes the experts' weights and activations to 4-bit precision, a format Chinese chips can process natively. In Meituan's load tests, throughput rises 43.3 percent for prompt processing and 42.77 percent for response generation compared with an 8-bit version. Meituan says accuracy stays on par with the uncompressed model.

The self-hosting guide describes a reference 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's no guidance for Nvidia GPUs at all.

Where does it land on the benchmarks?

The LongCat team reports 88.4 points on Terminal Bench v2.1, which measures how well agents solve command-line tasks. Meituan's comparison table lists Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8, and Deepseek-V4.1-Flash at the top with 90.6.

On DeepSWE, which tests how well models fix bugs in real code repositories, LongCat-2.5 jumps from 9.7 to 64.6 points. That places it behind every other model in the table except Claude Opus 4.8, with Deepseek-V4.1-Flash leading at 74.2 and GPT-5.6 Sol at 72.7. The more recent Claude Opus 5.5 and GPT-6.1 Sol are not included in the table.

The harder Terminal Bench v4.0 puts LongCat-2.5 at 38.5 points — slightly above GPT-5.6 Sol's 37.3 but well below Opus 5's 53.9. One bright spot: the model posts 91.9 on the web search benchmark BrowseComp, the highest score in the table, though Meituan ran that test in its own environment, which is worth keeping in mind when comparing numbers.

The takeaway: LongCat-2.5 sits at the level US models like Claude Opus 5 reached a few months ago on coding benchmarks. That's not frontier leadership, but it's close enough that the gap is now measured in months, not years.

What does it cost and how can you use it?

The weights are available under the MIT license, which allows commercial use, modification, and redistribution. That's among the most permissive licenses in open-weight AI.

Users can also try the model in a chat interface at longcat.ai. The API supports both the OpenAI and Anthropic formats and plugs into agent harnesses like Claude Code, Codex, and Hermes. At launch, Meituan is charging a limited-time promotional price of roughly €0.27 per million input tokens and €1.07 per million output tokens.

Why does this matter for the broader market?

Chinese open-weight models like Kimi K3.1 and GLM-5.3 have largely caught up with top US models on common benchmarks. LongCat-2.5 extends that trend into multimodal agent models, and it adds a hardware dimension: models optimized for Chinese chips from the start make that hardware more attractive outside China too.

Meituan isn't stopping at agents. The company previously released LongCat-Image, an open-source text-to-image model that gets by with just 6 billion parameters and prioritizes efficiency over sheer model size, and it continues to work on image generators. An open, Chinese-trained agent model that runs natively on Huawei silicon gives buyers outside the US-Chinese duopoly a third path — and gives both Nvidia and closed US model labs a new reason to watch the benchmark tables.

By Sophie Lindqvist

Original: huggingface.co

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

625 articles

Related articles

  1. Nvidia launches open-weight Nemotron models for enterprise agents and AI PCs
  2. Smuggled Chips and Server Loopholes: How China Gets Nvidia GPUs
  3. GPT-6.1 Terra halves API prices and closes the gap on OpenAI's biggest models
  4. NVIDIA Ships Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
  5. NVIDIA to invest up to $100 billion in OpenAI for 10 gigawatts of AI infrastructure

« Previous article