Meituan Open-Sources LongCat-2.0: China's 1.6T AI Model Built Without NVIDIA

Meituan, China’s food-delivery and tech conglomerate, just open-sourced LongCat-2.0 — a 1.6-trillion-parameter Mixture-of-Experts model that represents the largest AI system ever pre-trained entirely on domestic Chinese chips. Released June 30 under an MIT license, it had already been quietly topping OpenRouter usage charts under the codename “Owl Alpha” before Meituan revealed its identity.

What makes this different

Most Chinese AI labs, including DeepSeek, have relied on domestic hardware only for inference — the cheaper, less compute-intensive step of running a pre-trained model. Pre-training a frontier-scale MoE at 1.6 trillion parameters is an orders-of-magnitude harder hardware challenge.

Meituan ran LongCat-2.0 pre-training across a cluster of over 50,000 domestic ASICs. No NVIDIA H100s, no H200s, no A100s. The company hasn’t identified the specific chip supplier, but the existence of the run itself is a significant engineering claim — one that will require independent verification, though OpenRouter usage patterns lend it some credibility.

Performance

On agentic coding benchmarks, LongCat-2.0 lands in the same bracket as GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.6. It is not beating the current frontier, but near-frontier performance on a model trained exclusively with domestic chips — open-sourced commercially under MIT — is the story here.

The model ships with a 1-million-token native context window. Mixture-of-Experts architecture means active parameter count per inference step is a fraction of the total 1.6T, keeping practical inference costs manageable despite the headline parameter count.

Geopolitical weight

The timing is sharp. The U.S. export controls on advanced AI chips have been tightening across 2025–2026. Every flagship Chinese model trained on Nvidia hardware was, implicitly, a dependency on a supply chain that Washington can cut. LongCat-2.0, if the domestic-chip training claim holds up, breaks that dependency for pre-training — historically considered the part you couldn’t do without western silicon.

DeepSeek proved you could deliver frontier-quality inference with efficient architectures. LongCat-2.0’s claim is stronger: it’s asserting you can build the model from scratch on Chinese hardware.

MIT license and practical access

The full model weights are available under MIT — no usage restrictions, commercial use permitted, no revenue caps. That puts it squarely in competition with Meta’s Llama family as an enterprise-deployable open-weight option, with the added provenance of demonstrating domestic chip viability.

This doesn’t solve China’s compute deficit overnight. 50,000 ASICs at the scale required for 1.6T parameters still represents enormous infrastructure investment. But as a proof-of-concept that domestic-chip pre-training is viable at frontier scale, LongCat-2.0 is a meaningful milestone.

Sources

• VentureBeat — Meituan open sources LongCat-2.0, the 1.6T near-frontier agentic coding model trained entirely on Chinese chips
• South China Morning Post — China claims biggest AI model trained on local chips
• Cryptopolitan — Meituan open-sources 1.6-trillion-parameter LongCat-2.0 AI model trained on Chinese chips