LongCat : Meituan's open-source trillion-parameter MoE LLM with 1M token context.

LongCat is Meituan's open-source MoE model - the first trillion-parameter LLM trained end-to-end on domestic Chinese hardware. It packs 1.6T total params, ~48B activated, and a 1M-token context. As Owl Alpha it ranked top-3 on OpenRouter, #1 on Hermes, #2 on Claude Code, competing with GPT-5.5 and Opus 4.8. Trained on 30T+ tokens. MIT-licensed, including weights, infra, and inference engine - fully self-hostable.

Is LongCat censored on the official platform?

The official longcat.chat website may apply content filters per Chinese regulations. However, the open-source MIT-licensed model weights have no built-in censorship — self-hosted versions operate without restrictions. This is a key distinction from API-only Chinese AI platforms.

How does LongCat compare to GPT-4 or DeepSeek?

LongCat-2.0 packs 1.6T parameters (MoE architecture) with a 1M-token context window. On OpenRouter it ranks top-3 globally(2026-07-01). It's especially strong on agentic tasks. Unlike GPT-4, it's fully open-source. Compared to DeepSeek, it has more parameters and was trained entirely on domestic Chinese hardware.

Can I access LongCat from the United States?

Yes. Both the official website longcat.chat and the open-source model on GitHub are fully accessible from the US without a VPN. No Chinese phone number is needed. The model's MIT license has no geographic restrictions.

Is LongCat open source and free to use?

Yes. LongCat is MIT-licensed and its model weights are publicly available on GitHub. The official chat website longcat.chat also offers free tier access. Developers can self-host the model without restrictions, making it one of the most accessible Chinese LLMs globally.