Meituan announces LongCat-2.0 trillion-parameter MoE model

Category: AI
Status: developing
Sources: LongCat blog, Hugging Face model page, VentureBeat, discussion
Summary: Meituan's LongCat team announced LongCat-2.0, a Mixture-of-Experts language model with about 1.6 trillion total parameters and roughly 48 billion activated per token (dynamic activation reported between 33 and 56 billion), a 1-million-token context window, and an MIT license. The blog and reporting state the model was pretrained on more than 35 trillion tokens entirely on AI ASIC superpods rather than GPUs, and that it served as the stealth "Owl Alpha" model on OpenRouter for the prior two months. The Hugging Face model page lists the MIT license but states the weights are "coming soon," so the open-weight checkpoint is not yet published. Benchmark and OpenRouter-ranking claims come from the vendor and secondary reporting and are not independently reproduced.
Why it matters: A trillion-parameter model with a 1M context and a permissive license, reported as trained on non-NVIDIA domestic accelerators, adds to open-weight competitive pressure and is a data point on training large models off the GPU stack, though the weights and capability claims are unconfirmed.
Follow-up: Watch for the actual weight release on Hugging Face, a model card with reproducible benchmarks, and independent agentic-coding evaluations.