Meta opens Llama 3.5 Base weights for design-focused fine-tuning

AI · 4 min read

Meta published Llama 3.5 Base weights and accompanying guidelines that make it easier for third parties to fine-tune models on domain-specific datasets. The company highlighted use cases in product copy, pattern naming, and design-system taxonomy generation.

Alongside the weights, Meta released a recommended fine-tuning recipe for low-shot adaptation to brand voice and a tokenization tip sheet to preserve designer-friendly labels like 'primary-button' and 'error-toast'. Meta emphasized safety and IP controls for organizations that reuse proprietary design data.

Several design tooling startups announced immediate plans for fine-tuned Llama 3.5 variants to provide offline assistants and enterprise on-prem options. The availability of the weights is expected to increase competition in the design-assistant space.