Meituan says its new AI model rivals Anthropic's recent best, built on Chinese chips. No one has checked yet

Meituan, the Chinese food-delivery and local-services giant, made LongCat-2.5 generally available through its API on November 2, about five weeks after a September 25 preview. Its size carries over from LongCat-2.0, which shipped in June: roughly 1.6 trillion parameters, with about 48 billion in use for each piece of text on average.

What has changed, Meituan says, is the model's job and the hardware story. It is built for long, multi-step agent work such as coding and visual design. The company again says training ran entirely on Chinese-made chips. It now adds that the model can run in a compressed 4-bit format on that same domestic hardware without losing quality.

Those claims matter to anyone tracking China's push to build frontier AI without Nvidia. They are also almost entirely self-reported. No independent lab has published results for LongCat-2.5. Meituan still has not named the chips. And the model files it says it will release on GitHub and Hugging Face were not out when the model went live.

What Meituan is actually claiming

The pitch centers on handling images as well as text, and on endurance. Meituan says visual input is converted into the same units the model uses for text. In training, it describes large-scale learning by trial and feedback on long task sequences, including a pipeline that turns failures its agents hit in real use into training tasks. Chinese business outlet EEO reported that mid-training used sequences as long as one million tokens.

The commercial angle is distribution. Meituan's API changelog advertises compatibility with Claude Code, OpenAI's Codex and other agent harnesses, the software that gives coding agents their tools. Meituan is positioning LongCat as a cheaper engine for workflows built around American models. Third-party tracker LLM Stats lists Meituan as the only provider, at $0.30 per million input tokens and $1.20 per million output tokens. None of the launch reports we reviewed gave pricing.

The public benchmark record is thinner than for most frontier launches. According to EEO, Meituan says LongCat-2.5 lands between Anthropic's Claude Opus 4.8 and Claude Opus 5 across several mainstream benchmarks, and beats Opus 5 on some. Neither TechNode's launch report nor the LLM Stats listing carried specific scores. A launch summary published by Magica noted, without naming them, that competing models still lead several of Meituan's own comparisons. A placement "between" two models, without the underlying table, is a marketing frame rather than a result.

The Chinese-chip claim, again

The chip story is why this release matters beyond Meituan. In June, the company said LongCat-2.0 had been trained on a cluster of more than 50,000 domestic custom AI chips. As RITS reported, it billed that model as the industry's first trillion-parameter model to be both trained and run entirely on alternative hardware. At the time, The Decoder noted that Meituan had not named the chipmaker, and that without a Hugging Face release, outside checks were hard. RITS said the claim would need independent scrutiny. Four months later, that gap is unchanged. EEO's launch-day report says Meituan worked with domestic chipmakers but identifies none, and gives no chip count for LongCat-2.5.

What Meituan has added is engineering detail. According to EEO, the team kept most data traffic inside a high-speed "super node," and in one long-text training test, output per chip rose 86% over a standard baseline. To run the model, Meituan compresses both its stored weights and its working values to 4 bits, a method called W4A4. Meituan says this matches the uncompressed model on the downstream tasks it evaluated. It also lifts throughput about 40% and cuts the wait for a first reply about 16% versus an 8-bit baseline, according to EEO.

If those numbers hold up, they address the real constraint on Chinese chips, usually described as weaker per chip and less mature in software than Nvidia's. Aggressive compression and smart splitting of work are exactly how you fit a 1.6-trillion-parameter model onto weaker hardware. But each figure is Meituan's own, set against a comparison point it picked, on chips it will not identify. Magica's summary likewise cautions that the efficiency gains rest on specific test baselines. Quality loss from 4-bit compression can also show up unevenly, in long reasoning chains and edge cases rather than headline averages. That is exactly where a long-task agent model is supposed to excel.

Why the evidence gap matters

Meituan does have a record of shipping. LongCat-2.0's model files are now on the company's Hugging Face page. RITS reported that it handled roughly 10 trillion tokens a month on OpenRouter under the codename "Owl Alpha" — real usage, not just a press cycle. That history earns LongCat-2.5 a hearing, not a pass. And the company is marketing comparisons with Anthropic's Opus models before outsiders can run the model or test it on their own tasks.

There is also a strategic tension. On Meituan's August 28 earnings call, CEO Wang Xing said the company "will not become a token factory," EEO reported. The launch emphasized business use: the company's Miaoshou management workbench, built on LongCat models, already serves chain brands such as Guming. Yet the clearest outside data point is a low usage price and plug-in compatibility with rival coding tools. That looks a lot like selling tokens. Magica's summary adds that nothing in the launch shows Miaoshou merchants running on LongCat-2.5 specifically, or what a completed business task actually costs.

What to watch

Three things would turn this from a claim into evidence. First, the promised Hugging Face and GitHub release, ideally with the 4-bit version, so outsiders can test whether quality holds on long agent runs. Second, independent scores on agent benchmarks such as SWE-bench Pro and Terminal-Bench, from evaluators using their own test setups rather than Meituan's charts. Third, any identification of the chip vendor. Without it, the most geopolitically interesting claim — that a trillion-parameter model can be trained and served at scale on Chinese silicon — rests on trust.

Until then, LongCat-2.5 is a strong statement of intent wrapped in numbers nobody outside Meituan has yet been able to check.