For years, the narrative was simple: if you want to train frontier AI, you need NVIDIA GPUs. China's export-restricted access to high-end chips was supposed to be a bottleneck that would keep its models a generation behind.
Meituan just made that argument obsolete.
On June 30, the Beijing-based food delivery and local services giant open-sourced LongCat-2.0 — a 1.6-trillion-parameter mixture-of-experts (MoE) large language model that was trained entirely on domestic Chinese semiconductors. No NVIDIA H100s. No Blackwells. Just homegrown hardware.
What Makes LongCat-2.0 a Milestone
- 1.6 trillion parameters — making it one of the largest open-weight models ever released
- 1 million token context window — matching the best frontier models from OpenAI and Anthropic
- Fully open-source — weights, training code, and inference pipeline available on GitHub
- Chinese chip-native — trained on domestic accelerators (likely Huawei Ascend 910C)
This isn't just another model dump. It's a statement of training closure — China's AI ecosystem has proven it can build world-class models without access to Western hardware. Industry analysts at Pandaily called it "the moment China's domestic AI supply chain achieved full independence."
Why Developers Should Care
LongCat-2.0 arrives with strong benchmarks across Chinese-language reasoning, mathematical problem-solving, and long-context retrieval. For developers building on open-source models, it offers a massive, capable alternative to Llama, DeepSeek, and Qwen — particularly for applications requiring deep Chinese language and cultural understanding.
The model was trained at Meituan's scale — the company processes tens of millions of daily deliveries and has massive real-world data pipelines. That operational know-how is baked into the model's architecture.
The Bigger Picture
LongCat-2.0 lands at a moment when the AI world is already drowning in new releases — OpenAI's GPT-oss family, Huawei's openPangu 2.0 Flash, Anthropic's Claude Fable 5 redeployment, and Kimi K2's SWE-Bench record. But this one matters most for geopolitical AI strategy. If China can now train frontier models on domestic silicon, the chip export ban loses its teeth.
The implications for supply chains, national AI sovereignty, and the balance of open-source power are enormous. LongCat-2.0 isn't just another model — it's the sound of a bottleneck shattering.
Check it out on GitHub and try the demos. The age of chip-independent AI training has begun.
Top comments (0)