• Sources: primary
  • Summary: Figma first fine-tuned a Llama 3.3 70B model on more than 250,000 synthetic edit examples, generated by teacher models and checked by deterministic validators and visual judges. It now post-trains newer bases such as Qwen 3.8 27B and adds RL with design-specific environments and expert-written reward criteria. The Figma agent routes precise edits to this model and other tasks to third-party models, and Figma reports tasks about 20% faster and 60% fewer undos.
  • Why it matters: It is a concrete case of a narrow post-trained model replacing frontier calls for a high-frequency, non-verifiable task, with the eval split by dimension.