Meituan open-sources LongCat-2.0, a 1.6T coding model trained on Chinese chips

Read the original article →

What happened

Meituan open-sourced LongCat-2.0 on June 30, releasing it through GitHub, Hugging Face and its own platform under an MIT license. The model is a 1.6-trillion-parameter Mixture-of-Experts system with a native 1-million-token context window and scores about 59.5% on SWE-Bench Pro.

Before the reveal, the model had been leading OpenRouter usage anonymously under the codename 'Owl Alpha.' Meituan says it was trained from scratch on a cluster of roughly 50,000 domestic Chinese AI ASICs rather than NVIDIA GPUs.

Why it matters

A permissively licensed, near-frontier coding model that anyone can download shifts leverage toward teams that want to self-host. The MIT license removes many of the usage restrictions attached to other open-weight releases.

If independently verified, the domestic-chip training claim may be the more significant part of this release. Training a model this large entirely on domestic chips suggests China's AI compute stack is maturing faster than export controls assumed.

MintedBrain take

The OpenRouter leaderboard run under a codename is a useful signal: developers rated it on output before knowing its origin. Benchmark it against your own tasks, and factor in the operational cost of hosting a 1.6T-parameter model.

References

This article was originally published at SiliconANGLE. For the full piece, read the original article.

Discussion

  • Loading…

← Back to News