Meituan's LongCat-2.0 Is a 1.6T Frontier Model That Never Touched an Nvidia GPU
A food-delivery company just open-sourced a 1.6-trillion-parameter frontier model, and the part that should make Jensen Huang uncomfortable is how it was trained: fifty thousand-plus Chinese-made AI ASICs, zero Nvidia.
LongCat-2.0 comes from Meituan, yes, the Chinese Meituan you order dinner from, and it's not a toy. 1.6T total parameters, roughly 48B active per token, 35 trillion training tokens, native one-million-token context, a homegrown LongCat Sparse Attention, and it had been quietly topping OpenRouter for agentic coding before anyone clocked what it was. Weights and inference code went fully MIT on July 5. It ships with day-one Claude Code, OpenClaw and Hermes integration, and the API undercuts everyone, thirty cents per million input tokens during the launch promo with cached reads free.
The technically remarkable claim is no rollbacks, no irrecoverable loss spikes across a 35-trillion-token run on fifty thousand domestic accelerators. Anyone who's babysat a large training run knows how absurd that is. Training a trillion-parameter model without a single catastrophic blowup, on chips that aren't the industry-standard silicon, is a systems achievement independent of the model quality.
Here's why it matters beyond the benchmarks. The entire US export-control strategy rests on the assumption that frontier training needs Nvidia. LongCat-2.0 is the loudest counterexample yet: a near-GPT-5.5-class open-weight agentic coder, built end-to-end on Chinese chips, given away for free with a permissive license. Whether or not it beats Claude on your eval, it changes the argument. The moat was supposed to be compute access. Meituan just poured a bucket of cold water on that.
Model: longcatai.org, MIT licensed
← Back to all articles
LongCat-2.0 comes from Meituan, yes, the Chinese Meituan you order dinner from, and it's not a toy. 1.6T total parameters, roughly 48B active per token, 35 trillion training tokens, native one-million-token context, a homegrown LongCat Sparse Attention, and it had been quietly topping OpenRouter for agentic coding before anyone clocked what it was. Weights and inference code went fully MIT on July 5. It ships with day-one Claude Code, OpenClaw and Hermes integration, and the API undercuts everyone, thirty cents per million input tokens during the launch promo with cached reads free.
The technically remarkable claim is no rollbacks, no irrecoverable loss spikes across a 35-trillion-token run on fifty thousand domestic accelerators. Anyone who's babysat a large training run knows how absurd that is. Training a trillion-parameter model without a single catastrophic blowup, on chips that aren't the industry-standard silicon, is a systems achievement independent of the model quality.
Here's why it matters beyond the benchmarks. The entire US export-control strategy rests on the assumption that frontier training needs Nvidia. LongCat-2.0 is the loudest counterexample yet: a near-GPT-5.5-class open-weight agentic coder, built end-to-end on Chinese chips, given away for free with a permissive license. Whether or not it beats Claude on your eval, it changes the argument. The moat was supposed to be compute access. Meituan just poured a bucket of cold water on that.
Model: longcatai.org, MIT licensed
Comments