News

Releases & Models

Meituan launches LongCat-2.5, a 1.6 trillion parameter model trained on Chinese chips

The Chinese delivery giant launched a 1.6 trillion parameter model trained entirely on domestic chips, with aggressive pricing and benchmarks placing it near Claude Opus 5 on agentic tasks.

Ilustracao editorial abstrata de uma rede neural dourada sobre um wafer de chip, representando o modelo LongCat-2.5 da Meituan treinado em chips chineses

Beijing proves it can train a giant model without American chips

Meituan, better known for food delivery in China, just launched a 1.6 trillion parameter language model called LongCat-2.5, trained entirely on Huawei Ascend accelerators. It is a direct signal to the market: the company does not need American GPUs to compete at the frontier of artificial intelligence. The model has been live on Meituan's own chat and API since November 2, five weeks after a preview build circulated among testers.

What LongCat-2.5 brings

The model uses a mixture of experts architecture, with only 48 billion parameters activated per token despite 1.6 trillion total, which keeps usage cost low. Meituan describes LongCat-2.5 as natively multimodal: instead of bolting a separate image reader onto a text model, it turns images into discrete tokens that flow through the same pipeline as words. In practice, that lets it turn a screenshot or a design sketch into working code, and the company is betting heavily on long running agentic tasks, like full coding projects, instead of isolated short answers. Training used sequences of up to 1 million tokens, and the company says it reached 43% higher inference throughput running aggressive 4 bit quantization, which cuts the cost of each call significantly.

How it compares with Western models

The numbers released by Meituan itself place LongCat-2.5 between Claude Opus 4.8 and Claude Opus 5 on most tests, and the company claims it beats Opus 5 on browser automation tasks. On Terminal Bench v2.1, which measures command line ability, LongCat-2.5 scored 88.4 against Opus 5's 89.1, a small gap. On the newer and harder Terminal Bench v4.0, the gap widens a lot: 38.5 against 53.9. Pricing stands out on its own: $0.30 per million input tokens and $1.20 per million output tokens, well below what frontier Western models charge.

What is still missing to confirm the leap

Every benchmark on that table came from Meituan's own lab, with no independent audit so far, and the company promised to publish the weights on GitHub and Hugging Face but has not delivered yet: the official page still marks both links as coming soon. That means for now LongCat-2.5 can only be used through the API or Meituan's own chat, with no way to run it locally or independently verify whether the numbers hold outside the tests the company chose. Until the numbers land on public leaderboards and the weights actually show up, the model counts as a real launch confirmed by Meituan itself, but with performance still unvalidated by third parties.

Sources

PosixAI: Meituan's LongCat-2.5 Takes On Claude Opus 5 With 1.6 Trillion Parameters; eesel: LongCat 2.5, Meituan's new multimodal coding model, explained; TestingCatalog: launch note on X.