LongCat-2.0 is a 1.6-trillion-parameter open-source mixture-of-experts AI model from Meituan, built for agentic coding, with a native 1-million-token context window and roughly 48 billion active parameters per token.

Was LongCat-2.0 really trained without Nvidia chips?

Meituan claims LongCat-2.0 was trained end-to-end — both pre-training and inference — on a 50,000-chip domestic ASIC cluster with no Nvidia GPUs involved, using the Huawei Collective Communication Library and Atlas-950 SuperPods.

Can I download and use LongCat-2.0 right now?

It's licensed under MIT, but the model weights are not yet posted — both GitHub and Hugging Face list them as 'coming soon.' Only the inference framework and infrastructure code are currently available.

How does LongCat-2.0 perform on coding benchmarks?

Meituan reports a score of 59.5 on SWE-bench Pro, ahead of GPT-5.5's 58.6 and above Claude Opus variants on the same test. These are vendor-reported numbers; independent verification is still pending.

Meituan's LongCat-2.0: 1.6T Params, Zero Nvidia Chips

China's food delivery giant just open-sourced a 1.6-trillion-parameter coding model trained entirely on domestic chips — zero Nvidia GPUs. Meituan's LongCat-2.0 dropped June 30, and it's already topped OpenRouter's usage charts under the codename "Owl Alpha."

Here's what's real and what's still "coming soon."

The Specs

LongCat-2.0 is a sparse mixture-of-experts model: 1.6 trillion total parameters, but only 33–56 billion active per token (averaging around 48 billion). It ships with a native 1-million-token context window — enough to load a full codebase in one pass.

Meituan says it scores 59.5 on SWE-bench Pro, edging out GPT-5.5's 58.6 and beating Claude Opus variants on the same benchmark. Take that with the usual caveat: it's Meituan's own testing, and independent verification is still pending.

The Chip Story

This is the headline that matters. Meituan says LongCat-2.0 was trained end-to-end — pre-training and inference — on a 50,000-chip domestic ASIC cluster, using the Huawei Collective Communication Library and Atlas-950 SuperPods. No US hardware anywhere in the pipeline.

That's different from DeepSeek-V4-pro, which used domestic chips for inference but still leaned on other hardware for training. If Meituan's claim holds up under scrutiny, it undercuts the core assumption behind US export controls: that cutting off advanced GPUs would bottleneck frontier-scale training in China.

This lines up with the direction we've seen from Huawei's Ascend chip push against Nvidia — Chinese firms aren't just building alternative silicon, they're closing the software-stack gap that made CUDA hard to displace in the first place.

The Catch

It's licensed under MIT — about as permissive as open source gets, letting you fold it into closed-source commercial products without obligation. But the actual model weights aren't posted yet. Both the GitHub and Hugging Face pages currently read "Model weights coming soon."

What is available: the inference framework and infrastructure code. Training data composition, training duration, and total cost remain undisclosed — critics are already calling it "pseudo-open-source" until the weights actually land.

Try It Now

You can access LongCat-2.0 today through OpenRouter or Meituan's own LongCat platform, and pricing undercuts GPT-5.5 and Claude Sonnet 5 significantly — with context-cache hits processed free and a limited-time "Token Pack" pricing tier on top.

If you're evaluating it for real work, wait for the weights before betting anything critical on it — and treat the benchmark numbers as a starting point, not a verdict.

Sources: VentureBeat, South China Morning Post