Figma Trains Specialized Model to Speed Up Design Editing

Designers spend their days making countless small, iterative edits—spacing adjustments, color changes, layout refinements—work that’s difficult to verify or handle with general-purpose frontier models. Figma tackled this by fine-tuning models on 250,000+ synthetic examples pairing design states with requested edits, using Figma’s native file format so changes could be applied directly. Even a Llama 3.3 70B model trained this way reached parity with leading frontier models.

As Figma expanded to newer base models, supervised fine-tuning alone proved limiting—imitation learning reduced generalization and required massive example sets. The team shifted toward reinforcement learning with custom reward models calibrated to designer preferences. Quality assessment combines structural validation with visual judges and expert designer review, evaluating design across multiple dimensions rather than treating it as a single metric. This granular approach revealed what each model succeeded at and where it still needed growth.

Post-rollout metrics showed 20% faster task execution, 62% success in expert-judged comparisons with the prior agent, 33% higher user satisfaction, and 60% fewer undone edits. Figma doubled the share of work handled by its proprietary model while cutting per-request costs. The results validate that specialized post-training can match frontier models on bounded, domain-specific problems.