Meituan Open Sources LongCat-2.0 Coding Model Trained on Chinese Chips
Meituan has open-sourced LongCat-2.0, a 1.6 trillion parameter coding model trained entirely on domestically manufactured Chinese chips. The release signals a shift toward reducing reliance on US semiconductor technology in large-scale AI training.
Key Takeaways
- 1## Model Release and Scale Meituan, China's largest food delivery and local services platform, has made LongCat-2.
- 20 publicly available as an open-source project.
- 3The model contains 1.
- 46 trillion parameters and was trained exclusively on Chinese-manufactured processors, avoiding dependency on US-made GPUs or silicon subject to export restrictions.
- 5## Strategic Context The release occurs against the backdrop of US export controls on advanced semiconductor sales to China, which have limited access to high-end Nvidia and AMD chips since 2023.
Model Release and Scale
Meituan, China's largest food delivery and local services platform, has made LongCat-2.0 publicly available as an open-source project. The model contains 1.6 trillion parameters and was trained exclusively on Chinese-manufactured processors, avoiding dependency on US-made GPUs or silicon subject to export restrictions.
Strategic Context
The release occurs against the backdrop of US export controls on advanced semiconductor sales to China, which have limited access to high-end Nvidia and AMD chips since 2023. By demonstrating that a large-scale coding model can be trained on domestic infrastructure, Meituan's move signals progress toward Chinese technology self-sufficiency in AI development. The open-source publication also expands the pool of developers able to build applications on Chinese chip architectures.
Implications for Global AI Development
Successful training runs on non-US chips could accelerate development of alternative AI hardware ecosystems and reduce the concentration of large-model training among companies with unrestricted access to leading US processors. However, the computational efficiency and real-world performance of models trained on Chinese silicon relative to US-manufactured alternatives remain subjects for independent evaluation.
Why It Matters
For Traders
Crypto projects building on Chinese infrastructure or serving Chinese markets may benefit from reduced compute costs if domestic chip availability improves, though this effect is indirect and long-term.
For Investors
Demonstration of viable large-model training on non-US chips weakens the moat of US semiconductor dominance in AI and signals a structural shift in global compute geography over 12-24 months.
For Builders
Developers targeting Chinese users or operating under export-restricted jurisdictions now have a documented path to train and deploy models without relying on US hardware, widening the addressable surface for local infrastructure stacks.
