Meituan Ships a Lighter, Sparser LongCat-Flash

Meituan has released LongCat-Flash-Lite-Sparse, a lighter and more sparsely activated variant of its LongCat-Flash family of large language models. The model is published on Hugging Face under a permissive MIT license, making it freely available for both research and commercial use.

Like its siblings, the model uses a mixture-of-experts (MoE) architecture, in which only a subset of the network's parameters activate for any given token. The "Lite" and "Sparse" labels signal an emphasis on efficiency: fewer active parameters per forward pass typically means lower memory pressure and faster inference, which matters for teams running models at scale rather than in a lab.

Why it matters

Meituan is better known as one of China's largest on-demand delivery platforms than as an AI lab, and its steady stream of open releases is a reminder that serious open-weight work now comes from beyond the usual research houses. The choice of an MIT license is notable, placing the model among the more permissively licensed open releases available today.

A few things stand out about this release:

• It targets text generation as its primary modality.
• It leans on a sparse MoE design aimed at efficient serving.
• It carries an MIT license with no additional usage restrictions.

Meituan has not published detailed parameter counts or context-length figures alongside the initial listing, so practitioners will want to check the model card for configuration details and any benchmark numbers before deploying.