China's Meituan Open-Sources 1.6 Trillion-Parameter AI Model Trained Entirely on Domestic Chips

Meituan, China's food delivery giant, released LongCat-2.0 — a 1.6 trillion-parameter Mixture-of-Experts model trained from scratch on a 50,000-chip cluster of Chinese-made processors. The model, released under MIT license, matches or exceeds GPT-5.5 and Gemini on key coding benchmarks and had already been topping OpenRouter usage charts anonymously under the codename 'Owl Alpha.'

Key Takeaways

Meituan, China's food delivery giant, released LongCat-2.0 — a 1.6 trillion-parameter Mixture-of-Experts model trained from scratch on a 50,000-chip cluster of Chinese-made processors.
The model, released under MIT license, matches or exceeds GPT-5.5 and Gemini on key coding benchmarks and had already been topping OpenRouter usage charts anonymously under the codename 'Owl Alpha.'

AI synthesis3 sources3

The Model That Topped OpenRouter Before Anyone Knew It Was Chinese

On June 30, 2026, Meituan — China's food delivery juggernaut better known as the DoorDash of the East — dropped a bombshell on the AI world. The company released LongCat-2.0, a 1.6 trillion-parameter Mixture-of-Experts large language model, under an MIT open-source license. The headline number alone would have been enough to make news. But the real story is where and how it was trained.

LongCat-2.0 was trained from scratch on a 50,000-chip cluster powered entirely by Chinese-made processors. Meituan claims it is the world's first trillion-parameter AI system trained and run exclusively on domestic hardware — a direct consequence of the US export control regime that has barred Chinese companies from accessing Nvidia's cutting-edge chips since 2022.

Benchmark Performance That Rivals the Frontier

Despite the hardware constraints, LongCat-2.0 delivers competitive performance. The model scored 59.5% on SWE-bench Pro, narrowly edging GPT-5.5 at 58.6%. On Terminal-Bench, it achieved 70.8%. Meituan says LongCat-2.0 matches or exceeds Google's Gemini 3.1 Pro and competes with Anthropic's Claude Opus on agentic coding tasks.

The architecture is a Mixture-of-Experts design with 1.6 trillion total parameters, activating an average of 48 billion per token (ranging from 33 to 56 billion depending on query complexity). It features a native 1-million-token context window, allowing it to process ultra-long documents.

The 'Owl Alpha' Story

Perhaps the most fascinating detail: LongCat-2.0 had already been lurking on OpenRouter for weeks under the anonymous codename 'Owl Alpha.' During that time, it became one of the three most-used models on the platform — developers voting with their API calls based purely on output quality, with no idea they were using a Chinese company's model. Meituan deliberately kept the attribution hidden to gather real-world feedback before facing the reputational risk associated with Chinese AI in the current geopolitical climate. By the time the reveal came, Owl Alpha had an established track record.

Geopolitical Earthquake

LongCat-2.0 is a proof of concept that China's domestic chip ecosystem — led by companies like Huawei and Enflame — can train frontier-scale models without US hardware. The MIT license means no regional restrictions and full permission to fine-tune and redistribute.

The timing is significant: the 19-day export control ban on Anthropic's Fable 5 in June 2026 accelerated Chinese investment in domestic chip-trained models. LongCat-2.0 is the most visible result so far.

What It Means

For global enterprises, LongCat-2.0 represents a genuine alternative in the AI stack. For US policymakers, it raises uncomfortable questions about whether export controls are accelerating the very outcome they were designed to prevent — a fully independent Chinese AI ecosystem. And for the AI market, it signals that the era of US-only frontier AI may be coming to an end.