Skip to main content

Meituan's LongCat-2.0 Goes Open Source: A Big Win for Domestic AI Chips

On July 6, Meituan dropped a bombshell in the AI world: it officially open-sourced its foundational large model, LongCat-2.0. This isn't just another model release—it's the first major large model trained and inferred entirely on domestic computing chips, a milestone that has the industry buzzing.

Trained on Homegrown Hardware

LongCat-2.0, which debuted on June 30, was trained on a massive cluster of 50,000 domestic chips. That's no small feat. Handling a trillion-parameter model on homegrown hardware proves that China's chip ecosystem can deliver both stability and high performance at scale. For years, the narrative around domestic chips has been one of catching up—this release suggests they've arrived.

Open Source, Open for Business

Meituan isn't keeping this under wraps. On July 6, the model weights, inference engine, and detailed technical docs went live on GitHub, Hugging Face, and ModelScope. The kicker? It's released under the MIT license, meaning anyone—from solo developers to large enterprises—can use it commercially for free. That's a game-changer for small and medium businesses that couldn't afford cutting-edge AI before.

Chip Makers Jump On Board

The response from the domestic chip community was swift. On the same day, three major players—Huawei Ascend, MoLeLineage, and Muxi Technologies—announced they've completed inference adaptation for LongCat-2.0. This isn't just about compatibility; it's a signal that the "domestic chips + domestic large models" ecosystem is maturing fast. When hardware and software align this quickly, the whole industry benefits.

What This Means for China's AI Future

This open-source move is more than a technical showcase. By building a complete ecosystem around LongCat-2.0, Meituan is laying the groundwork for self-reliance in AI. As more developers dive in, optimize, and build applications, the voice of domestic AI on the global stage is only going to get louder. For now, the message is clear: homegrown chips can power world-class models, and the open-source community is ready to run with them.

Key Points

  • First domestic-chip-trained large model: LongCat-2.0 was trained and inferred on a 50,000-card domestic cluster.
  • Fully open source: Model weights, inference engine, and docs released under MIT license on GitHub, Hugging Face, and ModelScope.
  • Ecosystem support: Huawei Ascend, MoLeLineage, and Muxi Technologies have already adapted their chips for inference.
  • Low barrier to entry: MIT license allows free commercial use, empowering SMEs.
  • Strategic impact: Strengthens China's AI self-reliance and global competitiveness.