Ant Group Releases Ling-3.0-Flash Model Weights for Code-Based Web Design
Ant Group has released the open-source weights under an MIT license for Ling-3.0-flash, a 124-billion parameter Mixture-of-Experts (MoE) model. The model has demonstrated the ability to generate complex, styled webpages across various design aesthetics using only code like CSS gradients and SVG paths, without relying on external images. This release provides developers with a powerful local model capable of high-fidelity UI/UX generation, reducing reliance on external image assets. Its open-source nature and unique architecture could accelerate the development of AI-driven web design tools. The model activates only 5.1 billion parameters per token (1/64 expert activation) and features a hybrid-linear attention mechanism stacking KDA and MLA layers. The official FP8 quantized weights require approximately 128GB of storage and support a native 256K context window.
## BACKGROUND
A Mixture-of-Experts (MoE) architecture routes inputs to specific specialized sub-networks ("experts"), allowing large models to run efficiently by only activating a fraction of their total parameters per token. Generating visual layouts purely through code (CSS, SVG, and typography) requires the model to have a deep spatial and aesthetic understanding of how code translates into visual rendering.