Meituan unveils LongCat-2.0 AI model trained entirely on domestic Chinese chips

Meituan LongCat-2.0 Emerges as China's Self-Sufficient AI Breakthrough

Chinese food delivery giant Meituan officially unveiled LongCat-2.0 on June 30, revealing what it claims is the world's first trillion-parameter large language model trained and run entirely on domestic Chinese chips 1

1

. The 1.6-trillion-parameter Mixture-of-Experts model was developed on a cluster of over 50,000 domestically produced Application-Specific Integrated Circuits (ASICs), marking a definitive milestone for China's AI industry as it reduces dependence on U.S. hardware following export controls imposed by Washington since 2022 2

2

. The release demonstrates that near-frontier AI models can be scaled successfully without relying on Nvidia GPUs that have powered much of the global generative AI frontier model training effort 2

2

.

The Owl Alpha Mystery Solved: Two Months of Anonymous Dominance

Before its official debut, LongCat-2.0 operated anonymously on OpenRouter under the alias "Owl Alpha," commanding global developer charts for two months 2

2

. During this stealth period, the model processed approximately 10.1 trillion monthly tokens, averaging 559 billion tokens per day, representing a 242% month-over-month explosion in volume that propelled it into the platform's global top three 2

2

. By the time Meituan stepped forward, the agentic coding model had secured first place on the Hermes Agent workspace, second place on Claude Code deployments, and third place across OpenClaw environments 4

4

. The open-source AI model is now available on GitHub, Hugging Face, and Meituan's native platform under a permissive MIT license 2

2

.

Technical Architecture and Competitive Performance

The model activates roughly 48 billion of its 1.6 trillion parameters per token, with that figure ranging between 33 billion and 56 billion depending on query complexity 4

4

. LongCat-2.0 features a native 1-million-token context window, enabling it to handle ultra-long documents and complex coding tasks 1

1

. At its core lies an aggressive optimization of Mixture-of-Experts sparsity through a "Zero-Compute Experts" framework, ensuring routine execution elements pass through lighter subnetworks and eliminating idle computational overhead 2

2

. Meituan claims the model matched or exceeded several leading proprietary models, including Google Gemini, OpenAI's GPT-5.5, and Anthropic's Claude Opus, on some coding and agent benchmarks 1

1

. On SWE-bench Pro, which scores how often a model resolves real GitHub issues, LongCat-2.0 hit 59.5, ahead of GPT-5.5's 58.6 and Gemini 3.1 Pro's 54.2 4

4

.

Aggressive Pricing Challenges Western Models

Commercial access introduces a highly aggressive pricing tier where all context-cache hits are processed completely free of charge 2

2

. Standard API access runs $0.75 per million input tokens and $2.95 per million output tokens, slashed to $0.30/$1.20 during the current launch promotion 4

4

. This undercuts GPT-5.5's $5/$30 per million tokens and Claude Sonnet 5's introductory $2/$10 rate, landing close to DeepSeek V4-Pro's permanent $0.435/$0.87 4

4

. Meituan also provides token packs of 1 billion tokens at around $60, making it particularly attractive for coders and heavy users 4

4

. These pricing strategies arrive as Washington pressures top-tier American labs to restrict access to their latest models, with OpenAI forced to limit access to GPT-5.6 models and Anthropic taking Claude Fable 5 and Mythos 5 entirely offline 2

2

.

Self-Sufficiency Focus Amid U.S. Export Restrictions

LongCat-2.0's reliance on domestic Chinese chips underscores the growing importance of self-sufficiency in China's AI industry, as DeepSeek, Alibaba, ByteDance, and other major players work to reduce dependence on U.S. chips following export controls imposed by Washington 1

1

. The pretraining run spanned more than 35 trillion tokens across the cluster of over 50,000 domestically produced accelerators, finishing with "no rollbacks or irrecoverable loss spikes" 4

4

. This stability claim matters given how often large training runs on unproven hardware stacks fail midway through 4

4

. Chipmakers including Huawei and Enflame have moved quickly to fill the gap left by U.S. chipmakers, gaining market share through supply deals with AI developers 1

1

. According to Bernstein estimates, Nvidia held around 40% of the market share in China for AI chips in 2025, roughly matched by Huawei, with predictions that Nvidia's share will fall by 8% this year 5

5

.

Meituan's Strategic Diversification Through AI Agents

Meituan, often compared to DoorDash, is a late entrant to China's crowded and well-funded AI sector where rivals include DeepSeek and ByteDance's Doubao 1

1

. The LongCat team, founded in 2023, only launched its first model late last year 1

1

. Meituan has used earlier versions to power in-app AI agents that recommend restaurants and hotels and complete tasks such as ordering food and booking rooms, part of an "agentic commerce" trend rival Alibaba has accelerated this year 1

1

. Amid weak consumer sentiment and diminishing margins, Meituan may be seeking to diversify revenue streams, highlighting the model's ability to build gaming websites and write novels 1

1

. The model's architecture is designed specifically for agentic coding, helping it handle real-world coding tasks more efficiently and reliably 1

1

. This successful deployment of alternative silicon signals a structural shift that could threaten Nvidia's dominance if Chinese conglomerates can consistently iterate trillion-parameter architectures using homegrown ASICs rather than general-purpose CUDA GPUs 2

2

.