Meituan’s Giant AI Model Now Runs on Chinese Chips Alone

Human Verified 0

Share to: , and

A cluster of AI chips in a dark server hall with a cat silhouette formed by glowing circuit lines

Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter AI model built for autonomous coding agents, under the MIT licence. The weights arrived a week after the model’s unveiling, and the release is the first test of whether a frontier agent model can thrive entirely outside Nvidia’s ecosystem.

What was released

LongCat-2.5 is a mixture-of-experts model, which means only a subset of its specialist subnetworks works on each token. Out of 1.6 trillion parameters, roughly 48 billion are active per token. Through the API it handles up to one million tokens and generates responses up to 128,000 tokens long.

The model understands text, images and video, and it lets users switch reasoning on and off. Unlike many multimodal models, it does not bolt on a separate image encoder; it converts images into tokens that it processes alongside text, an approach aimed at visual coding, design and 3D tasks.

The hardware story

The hardware is the headline. The Decoder reports that LongCat-2.5 was trained entirely on Chinese hardware, and Meituan’s own posts describe domestic accelerators. The self-hosting guide documents a 194-chip Huawei Ascend Atlas A2 cluster, with no Nvidia guidance anywhere in it.

Meituan squeezes the model onto those chips with 4-bit quantisation, a format Chinese accelerators process natively. In its load tests, throughput rises 43.3 percent for prompt processing and 42.77 percent for response generation against an 8-bit build, with accuracy holding on par with the uncompressed model.

How it scores

On Terminal Bench v2.1, which tests command-line agent work, LongCat-2.5 scores 88.4. Meituan 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, a bug-fixing test, it jumps from its predecessor’s 9.7 to 64.6.

The harder Terminal Bench v4.0 puts LongCat-2.5 at 38.5, 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 score in Meituan’s table — though the company ran that test in its own environment.

Every score here is Meituan’s own, and no independent lab has published a result yet. The Decoder notes that the newer Claude Opus 5.5 and GPT-6.1 Sol are absent from Meituan’s comparison table, so the rankings say as much about the table as about the model.

Our read

Open weights change what this release means. A closed model trained on Chinese chips would be a press release; a downloadable one is a tool that labs, universities and startups can inspect, run and build on. That is why the weights, not the benchmarks, are the news.

The limits are just as real. Self-hosting is documented only on Huawei hardware, and the deployment needs 194 accelerators — far beyond a single server. Chinese models have largely caught up on benchmarks, but running in production is another matter.

The export-control angle is the one to watch. If a frontier-class model can be trained and served entirely on domestic Chinese hardware, chip restrictions lose leverage. The December hearings on Nvidia’s China business will now have a concrete example to argue over, as the Blackwell fingerprint findings already showed.

Why it matters

For the industry, the release raises the bar for open-weight models. Kimi’s K3.1 Pro and other Chinese systems have already narrowed the gap with US labs, and LongCat-2.5 adds a native multimodal agent to that list. The open-weight race is no longer a sideshow to the closed-model one.

Background

Meituan unveiled LongCat-2.5 on November 2 as a native multimodal model for long-horizon agent work, with the weights promised to follow. The weights arrived on November 10, a week later, in two Hugging Face repositories: the full model and a 4-bit quantised build.

Source: The Decoder report and LongCat-2.5 model card

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.