Meituan, the Chinese delivery app giant, has open-sourced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts (MoE) coding model that has secretly powered the top-ranked anonymous model on OpenRouter for two months. The system offers a native 1-million-token context window and targets autonomous software engineering, challenging closed-source enterprise incumbents.
The model is released under a permissive MIT license on GitHub and Hugging Face. For commercial API access, Meituan has introduced a tiered pricing structure that makes context-cache hits completely free. Uncached hits are priced at $0.75 per million input tokens and $2.95 per million output tokens—but a limited-time promotion drops those rates to $0.30 and $1.20, respectively.
This development is remarkable because LongCat-2.0 was trained entirely on Chinese-manufactured chips, signaling China's growing ability to produce frontier AI models despite Western export restrictions. The model's silent dominance on OpenRouter as "Owl Alpha" demonstrates its competitive performance against leading coding agents from US and European firms.
The MIT license and aggressive pricing pressure established players like GitHub Copilot and Cursor, which rely on closed-source models and per-seat subscriptions. Meituan's move could reshape the open-source coding agent market, forcing incumbents to either lower prices or accelerate their own open-source releases.
Meituan's foray into AI coding tools marks a strategic pivot from its core food delivery business. The company has not disclosed which Chinese hardware partner produced the training chips, but the achievement validates domestic semiconductor capabilities in high-performance AI workloads.
Counter-argument: Critics caution that LongCat-2.0's true performance outside OpenRouter's benchmark dashboards remains unverified, and enterprise adoption may be hindered by concerns over data sovereignty and geopolitical tensions. The model's open-source availability could also expose Chinese technology to scrutiny of potential biases or censorship embedded during training.
AI context: This brief is based on a single VentureBeat source published hours before the brief was composed. All numerical pricing, parameter counts, and model names are directly cited from the article. No independent verification of the model's OpenRouter performance or Chinese chip provenance has been conducted; claims are presented as reported by one outlet.