Meituan Open-Sources LongCat-2.0 Model Trained Entirely on Chinese AI Chips

Chinese technology company Meituan has open-sourced LongCat-2.0, a 1.6 trillion-parameter mixture-of-experts (MoE) large language model. The company says the model was trained and deployed entirely on domestically developed AI chips, underscoring the country's efforts to build AI infrastructure independent of NVIDIA hardware.

LongCat-2.0 is available to download on GitHub and Hugging Face, along with API access for developers.

According to Meituan, LongCat-2.0 matches or exceeds proprietary models such as GPT-5.5, Gemini 3.1 Pro and Claude Opus on several coding and agent benchmarks. These claims have not yet been independently verified.

“Both the full training run and the large-scale deployment are built entirely on AI ASIC superpods,” the company announced. Pretraining spans millions of accelerator-days across more than 35 trillion tokens, with no rollbacks or irrecoverable loss spikes. 

The release comes as Chinese AI companies increasingly shift toward homegrown chips following US export restrictions on advanced AI chips. While companies such as DeepSeek have previously used domestic hardware primarily for inference, Meituan claims LongCat-2.0 is the first trillion-parameter model to complete both pre-training and inference entirely on Chinese AI hardware.

It further said LongCat-2.0 activates around 48 billion parameters per token and was pretrained on more than 35 trillion tokens across millions of accelerator-days using AI ASIC superpods.

LongCat-2.0 supports a native one-million-token context window and introduces LongCat Sparse Attention (LSA), an updated sparse attention mechanism that combines streaming-aware, cross-layer and hierarchical indexing to reduce long-context inference costs. 

The model also incorporates an N-gram Embedding module, which expands the embedding space by nearly 100 times while accounting for less than 10% of the model's total parameters.

To support the model’s scale, Meituan said it trained LongCat-2.0 on more than 50,000 AI ASICs using a six-dimensional parallelism strategy, deterministic training operators, memory optimisations, and fault-recovery systems. The company said these optimisations improved training throughput by over 35% compared with a baseline implementation.

Meituan is positioning LongCat-2.0 as an agentic coding model. It said the model integrates with developer tools, including Claude Code, OpenClaw, and Hermes, for repository-level code editing, automated task execution, and software engineering workflows. The company also introduced three post-training expert groups focused on reasoning, agent execution and interaction quality.