Meituan releases LongCat-2.5 weights after training on Chinese chips
Tech

Meituan releases LongCat-2.5 weights after training on Chinese chips

TechNews Editorial
TechNews EditorialNov 15, 2026 · 3 min read
Share

Why it matters

Meituan's release offers a commercially usable agent model built for Chinese chips that, in its tests, approaches leading US models on some tasks.

The facts

  • Meituan released LongCat-2.5 weights under the MIT license one week after unveiling the agent model.
  • The 1.6-trillion-parameter model was trained entirely on Chinese hardware and uses about 48 billion parameters per token.
  • In Meituan's tests, it scored 88.4 on Terminal Bench v2.1 and 38.5 on Terminal Bench v4.0.

Advertisement

Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents. The release comes one week after its official unveiling. The Chinese food delivery company says the model, trained entirely on Chinese hardware, approaches top Western models on several benchmarks.

LongCat-2.5 is designed for autonomous, multi-step work, especially coding. It uses a mixture-of-experts architecture: roughly 48 billion of its 1.6 trillion parameters are active for each token. Through its API, it accepts up to one million tokens of context and can generate responses up to 128,000 tokens long. Users can switch reasoning on or off.

The model understands text, images and video, and responds in text. Building on LongCat-2.0, it adds image understanding by turning images at their native resolution into tokens processed alongside text. The approach draws on the LongCat-Next research model and aims to improve visual coding, design and 3D tasks.

For reinforcement learning, Meituan used software tasks drawn from GitHub pull requests and recreated failures by AI agents. Its training environments blocked internet access and removed Git history and cached files. A language model checked solutions for shortcuts. The team also distilled capabilities from specialized models for agent work, reasoning and dialog.

Read nextMicrosoft research shows how poisoned training can hijack coding agents

Meituan describes a Huawei chip deployment

Meituan offers a separate 4-bit version whose expert weights and activations can be processed natively by Chinese chips. In its load tests, that version increased prompt-processing throughput by 43.3 percent and response-generation throughput by 42.77 percent compared with an 8-bit version. Meituan says accuracy remained on par with the uncompressed model.

The self-hosting guide describes a deployment using 194 Huawei Ascend accelerators from the Atlas A2 series. Two handle images, 64 ingest prompts and 128 generate responses. The guide gives no instructions for Nvidia GPUs.

Two technicians push a fully loaded computing cabinet through a loading doorway into a machine room.
Illustration: AI & Tech News

Meituan reports mixed benchmark results

On Terminal Bench v2.1, which measures command-line task performance, Meituan reports 88.4 points for LongCat-2.5. Its table lists Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8 and Deepseek-V4.1-Flash at 90.6. That puts LongCat-2.5 close to the listed US models on this test.

On DeepSWE, a test of bug fixing in real code repositories, LongCat's score rises from 9.7 to 64.6. It trails every other model in Meituan's table except Claude Opus 4.8. Deepseek-V4.1-Flash leads with 74.2, followed by GPT-5.6 Sol at 72.7. The table does not include the newer Claude Opus 5.5 or GPT-6.1 Sol.

LongCat-2.5 scores 38.5 on the harder Terminal Bench v4.0, above GPT-5.6 Sol's 37.3 but below Opus 5's 53.9. It scores 91.9 on the web-search benchmark BrowseComp, the highest result in Meituan's table. Meituan ran that test in its own environment.

Meituan says it trained LongCat-2.0 in late June on more than 50,000 Chinese AI chips. That model previously appeared anonymously on OpenRouter as “Owl Alpha.” The company has also released LongCat-Image, an open-source text-to-image model with 6 billion parameters that prioritizes efficiency. Chinese open-weight models including Kimi K3.1 and GLM-5.3 have largely caught up with top US models on common benchmarks. Models built for Chinese chips could also make that hardware more attractive outside China.

The LongCat-2.5 weights are available under the MIT license, which permits commercial use, modification and redistribution. Its API supports OpenAI and Anthropic formats and works with agent harnesses including Claude Code, Codex and Hermes. Meituan's limited-time promotional price is roughly €0.27 per million input tokens and €1.07 per million output tokens. Users can now download the weights or try the model at longcat.ai.

Newsletter

Get the best AI & tech news daily

A concise daily digest. Unsubscribe anytime.

We use your email only to send this newsletter.

Keep reading