Meituan’s LongCat-2.0 is the first frontier AI model trained entirely on Chinese chips

Meituan, the Chinese food-delivery and commerce platform, released LongCat-2.0 to the public on July 3, 2026, under an MIT open-source licence — posting the model weights and training documentation to its public repository. The announcement, covered in detail by VentureBeat, disclosed that the model was trained on a cluster of 50,000 domestic Chinese AI accelerators, described as Huawei Ascend-class chips, with no Nvidia hardware involved at any stage of pre-training or inference.

What makes this different

Previous Chinese frontier models — including DeepSeek V4 Pro — were trained using domestic chips primarily for inference, the step of running a trained model to generate outputs. Pre-training, which requires sustained high-throughput matrix multiplication across enormous datasets, has been widely considered too demanding for current Chinese silicon to handle at frontier scale. LongCat-2.0’s disclosed training configuration directly challenges that assumption.

The model uses a sparse Mixture-of-Experts architecture with 1.6 trillion total parameters and a one-million-token context window. Its stated focus is agentic coding tasks — writing and executing multi-step software instructions — an area where it reportedly outpaced other public models on OpenRouter’s comparative leaderboard in the weeks before its public release. The 59.5% score on SWE-bench Pro, a standardised coding benchmark, places it within the range of models from major US labs, though direct comparison depends on test conditions and may shift as evaluation methods evolve.

Export controls and their limits

The US government has progressively tightened restrictions on AI chip exports to China since October 2022, targeting Nvidia’s A100, H100, and H200 lines and their successors. The stated rationale was that frontier AI capability requires chips that Chinese manufacturers cannot yet build domestically. LongCat-2.0’s training disclosure complicates that rationale — though analysts caution that it is difficult to independently verify training configuration claims, and that the model’s performance, while strong, may not yet match the frontier as defined by closed systems like GPT-5.5 or Claude Fable 5.

The disclosure arrives as Congress is weighing legislation that would subject AI chip exports to weapons-sale-level review, treating advanced semiconductors as a category of strategic military technology rather than a commercial product subject to case-by-case licensing. The bill’s supporters argue that events like LongCat-2.0 validate a stricter approach; its critics argue that the export controls have simply incentivised China to build domestic capacity it would otherwise have purchased from US firms.

What Meituan gains

Meituan’s decision to open-source the model under MIT rather than release it as a proprietary API marks a departure from the approach of most frontier labs. Open-sourcing at this scale makes the model freely usable for commercial applications, generating ecosystem engagement that can inform future development without requiring Meituan to run inference at scale. The company’s core business — logistics, food delivery, and local commerce in China — does not directly monetise LLM APIs, making a research-credibility play more valuable than a product-revenue play at this stage.

The model’s release follows a pattern established by DeepSeek and Alibaba’s Qwen series, in which Chinese labs have periodically released open-source models that close the gap with proprietary Western systems and force a recalculation of the assumptions behind AI policy. Whether LongCat-2.0’s training-hardware claims hold up to independent scrutiny will determine how significantly the release reshapes that debate.