Meituan Open Sources LongCat 2.0: 1.6 Trillion Parameter MoE Model Trained on AI ASICs

news

Meituan releases LongCat 2.0, a 1.6 trillion parameter MoE language model with 1 million token context window, trained entirely on AI ASIC hardware. Model weights coming soon under MIT license.

On June 30, 2026, Meituan (美团) unveiled LongCat 2.0, a massive Mixture-of-Experts language model with 1.6 trillion total parameters and approximately 48 billion activated parameters per token. The model is licensed under MIT and represents one of the largest open-weight models ever released.

Model weights are not yet available — the team notes "weights coming soon" on the HuggingFace repository. This article covers the announcement and architecture.

Key Specifications

SpecificationDetail
ArchitectureMixture-of-Experts (MoE)
Total Parameters1.6 trillion
Activated Parameters~48B per token
Context Window1 million tokens (LongCat Sparse Attention)
Training Data35+ trillion tokens
Training HardwareAI ASIC superpods (not NVIDIA GPUs)
LicenseMIT

Significance: AI ASIC Training

One of the most notable aspects of LongCat 2.0 is that both the full training run and large-scale deployment were built entirely on AI ASIC superpods — custom AI accelerator chips rather than NVIDIA GPUs. The pretraining spanned millions of accelerator-hours across more than 35 trillion tokens with no rollbacks or irrecoverable loss spikes, demonstrating frontier-scale training capability on alternative hardware.

Architecture Highlights

LongCat 2.0 introduces LongCat Sparse Attention, designed to handle long-context tasks efficiently. The model was trained on hundreds of billions of tokens of 1M-context data. Combined with dedicated post-training, this gives LongCat 2.0 strong performance on coding and agentic tasks.

Status

The HuggingFace repository (meituan-longcat/LongCat-2.0) is live with documentation and specifications, but model weights have not been released yet. The team has indicated weights will follow.

Meituan Open Sources LongCat 2.0: 1.6 Trillion Parameter MoE Model Trained on AI ASICs | ComfyUI Wiki