Meituan Open-Sources LongCat-2.0, Trained Entirely on Chinese Computing Infrastructure
Claire Weston
Meituan open-sourced its 1.6-trillion-parameter model LongCat-2.0 on July 6, trained and deployed end-to-end on domestically designed Chinese chips — not a benchmark play, but proof that China's homegrown compute stack can close the full loop on ultra-large models.
How big is this model, really?
LongCat-2.0 uses a Mixture-of-Experts architecture — the model is split into many specialist sub-modules, and only a fraction activate per query. Total parameters: 1.6 trillion; active per token: roughly 48 billion.
This means → the model is enormous on paper, but its real-time compute load is a fraction of the full size. That gap is exactly what lets it run on domestic chips.
It natively supports a 1-million-token context window, was pre-trained on over 30 trillion tokens, and peaked at more than 50,000 domestically designed accelerator cards during training.
How did an anonymous model crack the global top three in two months?
LongCat-2.0 first appeared on OpenRouter under the alias "Owl Alpha" in late April — no branding, no launch event, no press.
Within two months it ranked among the top three globally by monthly API calls: 10.1 trillion tokens processed per month, 559 billion per day, with a 242% month-on-month growth rate.
Users spanned the US, India, Kenya, Germany, and Spain, running code generation, tool-calling, and agent workflows. On June 27 an overseas blogger traced Owl Alpha back to Meituan; on June 30 Meituan confirmed the link.
Why were three Chinese chip makers ready on day one?
Huawei Ascend, Moore Threads, and MetaX all completed inference adaptation for LongCat-2.0 on the same day it was open-sourced.
In plain terms = all three had done the groundwork in advance. The model went live and their chips could run it immediately — not a coincidence, but a sign that the "domestic chips + domestic models" strategy has reached an executable stage.
Huawei's Ascend CANN team used prefill-decode disaggregation — splitting "understanding the input" and "generating the output" across different hardware — to fit a 1.6-trillion-parameter model onto cards with just 64 GB of memory each.
What role does the MoE architecture play here?
At inference, only about 48 billion parameters activate — not the full 1.6 trillion. This means → memory and compute demands drop sharply.
This reflects something bigger than an efficiency trick: MoE is the architectural precondition for running ultra-large models on Chinese chips. It routes around the "chips aren't strong enough" problem by design — each query uses only a small slice of the whole.
Is the performance good enough? How far behind the leaders?
Meituan did not position LongCat-2.0 as the world's best. On some benchmarks it still trails GPT-5.5 and Claude Opus 4.8.
Community feedback shows it approaches Claude Opus 4.6 on agent tasks — tool-calling, complex instruction execution — but has not yet matched Claude Opus 4.8.
In plain terms = performance is not top-tier, but it has entered the usable range. The point is not the ranking — it is that this usable level was achieved entirely on domestic compute.
What does this mean for the market?
The model is available on GitHub and Hugging Face under an MIT license — BF16 raw weights, FP8 and INT8 quantized versions, free for commercial use. Open-sourcing came just one week after the official release.
This means → Meituan chose the most permissive license available. Anyone can deploy it commercially, which will accelerate downstream validation of the domestic compute ecosystem.
This reflects a shift in how the market prices China's domestic compute supply chain over the long term: once a trillion-parameter model's full-stack training and commercial deployment have been completed on Chinese chips, the question "can domestic compute actually work?" moves from a technical debate to a delivered fact.
Content is for reference only, not financial advice.