Meituan open-sources 1.6T model: trained entirely on domestic chips, zero NVIDIA
July 2, Meituan open-sourced LongCat-2.0.
1.6T parameter MoE, 59.5 on SWE-bench Pro, pricing at $0.038 per million tokens — those numbers are eye-catching enough. But what really set the industry buzzing is something else: this model was trained end-to-end on domestic Chinese chips, without a single NVIDIA GPU.
First, the "Owl Alpha" disguise
LongCat-2.0 had been running on OpenRouter under the name "Owl Alpha" for quite a while — and ranked among the highest by call volume.
Meaning: many developers were already using Meituan's model without knowing it — and found it good enough to push it up the charts.
What does that tell you? Performance speaks. Not brand, not funding narrative — pure usage pushed it up.
Now the identity is revealed: Meituan. The food-delivery company quietly pulled off something big in AI.
Key stats
- 1.6T parameter MoE architecture
- Trained entirely on domestic ASIC chips, no NVIDIA GPUs
- SWE-bench Pro: 59.5 (coding benchmark, competitive with many hundred-billion-parameter closed models)
- Pricing $0.038 per million tokens, cache hits free
- Open-source release
What does $0.038 per million tokens mean? GPT-5.6 is $5/$30. LongCat is over a hundred times cheaper. Different capability tier, of course — but at this price, you can run it for workloads that don't need a frontier model, at nearly negligible cost.
Why "domestic chips only" is a big deal
This signal matters more than the model itself.
For two years, everyone's been asking one question: with NVIDIA export restrictions, can domestic chips train large models?
The answer was always "yes, but how good?" — domestic-chip-trained models were either small or visibly behind NVIDIA-trained ones.
LongCat-2.0 is the first case proving that pure domestic chips can train a 1.6T-scale, ~60-point SWE-bench Pro frontier model.
Not saying domestic chips match NVIDIA. Saying "good enough" — on the path to training frontier models, domestic chips can clear the bar.
What it means for the industry
- For Chinese AI firms: the compute-strangulation anxiety eases a bit. HBM and advanced nodes remain bottlenecks, but "can train big models" is now validated
- For NVIDIA: no short-term impact, but long-term, if domestic chips push training costs down, NVIDIA's pricing power gets squeezed
- For developers: one more ultra-cheap option. At $0.038, many AI applications that were too expensive to run become viable
Meituan's play
Meituan does food delivery and local services. Why train a large model?
Two reasons: first, its own business needs it — customer service, rider dispatch, merchant content generation all need AI. Second, Meituan doesn't want to be a pure application layer in the AI era; it wants infrastructure capability.
Open-sourcing LongCat-2.0 tells everyone: Meituan isn't just a delivery company. It has real substance on AI infrastructure.
The real value of LongCat-2.0 isn't how it stacks against GPT-5.6 — it won't beat that. The value is proof: domestic-chip-trained frontier models can run, work, and stay cheap. That's a structural shift in the AI compute landscape.