- Published on
Meituan open sources LongCat-2.0, the 1.6T, near-frontier agentic coding model that's been leading OpenRouter — trained entirely on Chinese chips
- Authors
- Name
- DeepHorus
Chinese delivery giant Meituan has officially open-sourced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts (MoE) model, under a permissive MIT license. The model, previously operating anonymously as 'Owl Alpha,' has dominated global developer charts on OpenRouter for two months, processing over 10 trillion monthly tokens. This release marks a significant shift in AI infrastructure, as LongCat-2.0 was trained entirely on a cluster of more than 50,000 domestic Chinese ASICs, bypassing reliance on Nvidia GPUs.
Developed for autonomous software engineering, LongCat-2.0 features a native 1-million-token context window and employs aggressive sparsity to activate only 33 billion to 56 billion parameters per token. During its anonymous tenure on OpenRouter, Owl Alpha accounted for approximately 10.1 trillion monthly tokens—averaging 559 billion tokens daily—a 242% month-over-month surge that propelled it into the platform’s global top three. The model also secured first place on the Hermes Agent workspace and second place on Claude Code deployments.
Meituan introduced a highly competitive pricing tier for commercial access: all context-cache hits are free, and a promotional discount reduces uncached input to $0.30 per million tokens and output to $1.20 per million tokens. Standard pay-as-you-go rates are $0.75/$2.95 per million tokens in/out. This pricing undercuts many top-performing models globally, making advanced AI more accessible.
The training on domestic ASICs rather than Nvidia hardware signals a profound structural shift in global tech infrastructure. It demonstrates that near-frontier AI models can be scaled successfully without the US-manufactured GPUs that have powered most generative AI training. This development threatens Nvidia’s dominance in the AI chip market and highlights China’s growing self-sufficiency in AI hardware.
The release coincides with increased US restrictions on Western AI models. Following government requests, OpenAI limited access to GPT-5.6, and Anthropic took its latest Claude models offline. Critics argue these measures have backfired, creating demand for affordable, open alternatives. LongCat-2.0 fills this gap, appealing to global developers seeking high-performance, cost-effective solutions.
Industry experts view LongCat-2.0 as a turning point in the AI landscape. Its 'Zero-Compute Experts' framework eliminates idle computational overhead, while dynamic activation ranges from 33 to 56 billion parameters depending on query complexity. This design enables efficient, scalable performance across diverse tasks.
Looking ahead, LongCat-2.0’s success with domestic chips suggests Chinese conglomerates can consistently iterate trillion-parameter architectures without relying on Western hardware. As geopolitical tensions rise and restrictions tighten, open-source models like LongCat-2.0 may reshape the global AI market, offering developers a viable path forward that sidesteps both hardware dependencies and costly proprietary APIs.
