Quick Facts
- LongCat-2.0 has 1.6 trillion total parameters and outscores GPT-5.5 on SWE-bench Pro with a 59.5 versus 58.6 rating.
- The model was trained on a 50,000-card domestic Chinese compute cluster, with no Nvidia or AMD chips used at any stage.
- Under the anonymous name “Owl Alpha,” the model ranked No. 1 on OpenRouter’s Hermes Agent chart and logged 10.1 trillion monthly tokens before its identity was revealed.
Meituan has open sourced LongCat-2.0, a 1.6 trillion parameter Mixture-of-Experts coding model that spent two months topping global developer charts under a secret alias. The company published the model on GitHub and Hugging Face on June 30, 2026.
The release confirms that “Owl Alpha,” an anonymous model that had been quietly tested on OpenRouter since April 24, 2026, was LongCat-2.0-Preview all along. In that time, it reached No. 1 on Hermes Agent, No. 2 on Claude Code, and No. 3 on OpenClaw, with 559 billion daily tokens and 242% monthly growth.
Performance Against Frontier Models
On SWE-bench Pro, LongCat-2.0 scores 59.5, beating GPT-5.5’s 58.6. It scores 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-bench Multilingual. Additional benchmarks include 73.2 on FORTE, 78.8 on RWSearch, and 79.9 on BrowseComp.
The model activates roughly 33 billion to 56 billion parameters per token despite its 1.6 trillion total count. Its context window reaches 1 million tokens, enough to process an entire large codebase in a single pass.
The Chinese Chip Story
Meituan claims LongCat-2.0 is China’s first trillion-parameter AI model trained entirely on domestically manufactured hardware. The training run spanned more than 35 trillion tokens across millions of accelerator-hours on a 50,000-card cluster, with no Nvidia A100s, H100s, or AMD MI300X chips involved.
Meituan did not name its hardware supplier. The company said in a WeChat post it used the Huawei Collective Communication Library to improve training stability. Community discussion on Hacker News and X points to Huawei Ascend 910C chips as the likely hardware.
The achievement sets LongCat-2.0 apart from DeepSeek’s flagship V4-pro model. DeepSeek used domestic chips only for inference. Meituan used them for both inference and the far more demanding pretraining process.
Architecture and Training
Three architectural features define LongCat-2.0. LongCat Sparse Attention handles long-horizon tasks and was trained on hundreds of billions of tokens of 1 million-context data. N-gram Embedding, inherited from LongCat-Flash-Lite, improves parameter efficiency with 135 billion N-gram Embedding parameters and an n-gram size of 5.
Post-training uses a method called Multi-Teacher Optimization via Mixture of Specialized Experts, or MOPD. This separates post-training optimization into three expert clusters: Agent, Reasoning, and Interaction. Each cluster is gate-routed per task rather than blended into a single reward function.
The full training run completed without rollbacks or irrecoverable loss spikes, which Meituan frames as proof that frontier-scale training on alternative hardware is viable.
Business Context
LongCat-2.0 is Meituan’s third major model release in under a year. LongCat-Flash launched in September 2025 at 560 billion parameters. LongCat-Next, a multimodal variant, followed in March 2026. Across the full LongCat series, Meituan has now released 12 models covering text, images, voice, and video.
Meituan CEO Wang Xing has tied the company’s AI push directly to its core food delivery and consumer services business. In a recent conference call, he said the company’s strategy is “proactive offence rather than passive defence” on AI, and committed to significantly increasing investment in AI agents and large model capabilities.
For U.S. software companies, the release signals that China’s chip independence efforts are advancing faster than many expected. A frontier-scale coding model trained entirely on domestic hardware, now open sourced and available for free on OpenRouter and Hugging Face, raises the competitive bar at no cost to developers who choose to use it.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.