Meituan, the Chinese food-delivery and services giant best known as the operator of the DoorDash-like Meituan Waimai platform, has open-sourced LongCat 2.0 — a 1.6-trillion-parameter mixture-of-experts language model trained entirely on Chinese-made AI ASIC chips, with zero Nvidia GPUs involved.
The model, released under the permissive MIT license on GitHub and Hugging Face, represents one of the largest open-weight models ever made available. LongCat 2.0 uses a mixture-of-experts architecture with approximately 48 billion activated parameters per token out of 1.6 trillion total, and supports a native context window of 1,000,000 tokens via a custom LongCat Sparse Attention mechanism.
Trained on 50,000 ASIC chips
The most significant aspect of the release is not the model’s size, but the hardware it was built on. Meituan trained LongCat 2.0 on a cluster of more than 50,000 Chinese AI accelerator chips — purpose-built ASICs rather than Nvidia GPUs. The pretraining ran across millions of accelerator-hours and more than 35 trillion tokens. The company reports no rollbacks or irrecoverable loss spikes during training, demonstrating that frontier-scale training is achievable on non-Nvidia hardware.
This matters because export controls have made it increasingly difficult for Chinese companies to access Nvidia’s most advanced GPUs. Meituan’s achievement suggests that the gap between Nvidia-powered and alternative-hardware training is narrowing faster than many in the industry expected.
Performance: beating GPT-5.5 on coding
LongCat 2.0 scored 59.5 on SWE-bench Pro, a repository-level software engineering benchmark, narrowly beating GPT-5.5’s 58.6. It was previously known as the anonymous “Owl Alpha” model on OpenRouter, where it ranked among the platform’s global top three models and topped the Hermes Agent Workspace leaderboard before its identity was revealed. At its peak, Owl Alpha processed approximately 10.1 trillion tokens per month, with 559 billion tokens daily on average and 242 percent month-on-month growth.
The model also introduces architectural innovations including Zero-Compute Experts and the MOPD post-training framework, both designed to improve long-context efficiency and agentic task performance.
MIT license
The MIT license is the most commercially permissive open-source option available. Developers can modify, redistribute, and integrate LongCat 2.0 into proprietary software without open-sourcing their derivative work. For enterprises, that means self-hosting the model to reduce API dependence, improve data privacy, and lower operational costs through free context-cache hits.
Model weights are listed as “coming soon” on the HuggingFace repository at the time of the announcement, with documentation and specifications already live.
Sources: Chinese DoorDash rival smashes open source record with 1.6-trillion-parameter LLM (TechRadar, July 9, 2026); Meituan Open Sources LongCat 2.0 (ComfyUI Wiki, June 30, 2026); Meituan Open Sources LongCat-2.0 Under MIT License (Open Source For You, June 2026)