Meituan's LongCat-2.0: A 1.6 Trillion Parameter Model Trained Entirely on Chinese Chips

Meituan's LongCat-2.0: A 1.6 Trillion Parameter Model Trained Entirely on Chinese Chips — and It's Open Source

Chinese delivery giant Meituan has open-sourced LongCat-2.0, a 1.6-trillion-parameter MoE model trained entirely on 50,000 domestic Chinese ASICs — no Nvidia hardware. It beats GPT-5.5 on SWE-bench Pro (59.5 vs 58.6), offers a 1M-token context window, and was secretly deployed on OpenRouter as 'Owl Alpha' where it became one of the top three most-used models globally.

Key Takeaways

Chinese delivery giant Meituan has open-sourced LongCat-2.0, a 1.6-trillion-parameter MoE model trained entirely on 50,000 domestic Chinese ASICs — no Nvidia hardware.
It beats GPT-5.5 on SWE-bench Pro (59.5 vs 58.6), offers a 1M-token context window, and was secretly deployed on OpenRouter as 'Owl Alpha' where it became one of the top three most-used models globally.

AI synthesis4 sources13

Chinese delivery and super-app giant Meituan has officially unveiled LongCat-2.0 on GitHub and Hugging Face, unmasking the model as the computational engine behind 'Owl Alpha' — the anonymous stealth model that spent the last two months commanding global developer charts on OpenRouter.

LongCat-2.0 is a 1.6-trillion-parameter Mixture-of-Experts (MoE) system with a native 1-million-token context window, released under the permissive MIT license. But what makes this a true inflection point is its operational independence: the massive model was trained entirely on a cluster of over 50,000 domestic Chinese Application-Specific Integrated Circuits (ASICs), proving that near-frontier AI can be scaled without Nvidia's GPUs.

Benchmark performance is striking. LongCat-2.0 scores 59.5 on SWE-bench Pro, surpassing OpenAI's proprietary GPT-5.5 at 58.6. It also achieves 70.8 on Terminal-Bench 2.1, 77.3 on SWE-bench Multilingual, and 73.2 on the FORTE enterprise workflow simulator. While it generally trails premium frontier systems like Claude Opus 4.8 on broad general benchmarks, it punches well above its weight in software engineering.

During its anonymous residency on OpenRouter as 'Owl Alpha', the model processed approximately 10.1 trillion monthly tokens — 559 billion per day — representing a 242% month-over-month explosion in volume that propelled it into the platform's global top three.

The reveal has strategic geopolitical implications. LongCat-2.0 arrives as Washington pressures American labs to restrict access to their latest models — OpenAI was forced to limit GPT-5.6, and Anthropic took Fable 5/Mythos 5 offline for 19 days. By locking down Western closed-source models and driving up API costs, critics argue the US has left a wide operational window for global developers seeking affordable, high-performance Chinese alternatives.

Commercial pricing is aggressive: standard API rates are $0.75/$2.95 per million tokens in/out, with a limited-time promo slashing uncached input to $0.30 and output to $1.20 — among the cheapest for top-performing models globally. Context-cache hits are processed completely free of charge, a game-changer for large-scale agentic coding workflows where repetitive repository reads would otherwise compound costs.

Meituan, best known as China's food delivery leader with 770 million annual transacting users and 14.5 million merchants, has pivoted aggressively into AI. This release follows LongCat-Flash (560 billion parameters) in late 2025 and positions the company as a foundational player in global AI infrastructure.