Meituan open-sources 1.6T coding model
LongCat-2.0, a 1.6-trillion-parameter MoE from Meituan, scores 59.5 on SWE-Bench Pro (ahead of GPT-5.5), MIT license, 1M token context.
Meituan, best known as China’s dominant food delivery platform, released LongCat-2.0 on June 30, a 1.6-trillion-parameter mixture-of-experts model trained entirely on around 50,000 domestic Chinese ASIC chips without any Nvidia hardware. The model carries an MIT license, supports a 1-million-token context window, and activates approximately 48 billion parameters per token.
On SWE-Bench Pro, LongCat-2.0 scores 59.5, placing it ahead of GPT-5.5’s 58.6 on that benchmark. The model was operating in the open under the name “Owl Alpha” and had spent roughly two months at the top of OpenRouter’s coding leaderboard before the announcement revealed its identity.
The weights are not downloadable yet. Both the GitHub and Hugging Face pages say “model weights coming soon.” The announcement is official and the benchmark numbers are cited by Meituan, but operators cannot run or evaluate the model locally until the weights publish.
What this signals
Two things are notable beyond the benchmark number. First, this is the largest model credibly claimed to have been trained without export-controlled Nvidia hardware, which matters for anyone tracking the practical gap between US and Chinese AI infrastructure. Second, Meituan is not an AI lab; it is an app company that built a frontier coding model as a side project, which says something about how accessible large-scale training is becoming.
MIT licensing removes commercial constraints once the weights land, making it directly comparable to other open-source coding models like Ornith 1.0 and the models tracked in the open models guide. Weights are worth monitoring; if they match the benchmarks, LongCat-2.0 becomes a serious contender for cost-sensitive agentic coding workloads.