Meituan open-sources LongCat-2.0, a 1.6-trillion-parameter MoE model trained entirely on Chinese chips

Meituan released LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts language model, on Hugging Face under the MIT license — the latest and largest entry in its LongCat family, and one it says was trained end-to-end on domestic Chinese accelerators rather than Nvidia hardware.

What's new

The model card lays out the scale directly: LongCat-2.0 is "a large-scale MoE language model with 1.6 trillion total parameters and ~48 billion activated per token." Other specifics from Meituan's own documentation:

• Training scale. "Pretraining spans millions of accelerator-days across more than 35 trillion tokens, with no rollbacks or irrecoverable loss spikes" — a claim aimed squarely at demonstrating training stability at frontier scale on non-Nvidia silicon.
• Long-context design. Meituan says it trained the model "on hundreds of billions of tokens of 1M-context data," pairing that with what it calls LongCat Sparse Attention to make million-token context practical at inference time.
• License and availability. "The model weights are released under the MIT License," with quantized INT8 and FP8 variants published alongside the full model, and support for both GPU and NPU deployment.

Context

LongCat-2.0 follows Meituan's earlier LongCat releases and arrives during a period of intense open-weight competition among Chinese labs — DeepSeek, Alibaba's Qwen team, and Moonshot's Kimi have all shipped large open models in recent weeks. What sets this release apart is the training-hardware claim: Meituan says both pretraining and serving ran entirely on domestic accelerators, at a scale (1.6T total parameters, 35T+ training tokens) that puts it in the same tier as leading Nvidia-trained frontier open models. Independent coverage has also connected LongCat-2.0 to "Owl Alpha," an anonymously-labeled model that had been topping developer usage charts on OpenRouter for weeks before Meituan revealed it as the same underlying system.

Why it matters

A model of this scale trained without Nvidia hardware is a direct data point in the ongoing question of how far US chip export controls have actually slowed Chinese frontier AI development — LongCat-2.0 suggests domestic accelerator stacks can now support training runs at a scale that was assumed to require Nvidia GPUs. Releasing the weights under a permissive MIT license, rather than a more restrictive research license, also lowers the bar for developers and enterprises anywhere to build directly on a near-frontier open model. Combined with the "Owl Alpha" reveal, it's a reminder that some of the strongest anonymous entries on public model leaderboards are turning out to be unannounced releases from major Chinese tech companies rather than experimental labs.