We're Training a Figma-Fluent Editing Model

Figma is developing a specialized model for the small, repeated edits designers make to existing canvas work, such as changing spacing, colors, or layouts in response to feedback. Rather than relying on extensive prompts to teach general-purpose models Figma’s file format, the team used supervised fine-tuning on synthetic examples pairing a design with an edit request and a successful change. Larger models generated candidate edits, which were checked for valid structure and whether they fulfilled the request; evaluator feedback could trigger revisions. An early Llama 3.3 70B model trained on more than 250,000 examples competed with leading models of its time. But imitation-based training can limit generalization and requires many examples, so the newer approach also uses reinforcement learning, with more varied tasks and rewards shaped by designer criteria.

Figma evaluates edits across separate dimensions, including instruction-following, structural correctness, visual quality, speed, and cost, combining automated checks and visual judges with expert review. In one internal comparison, an unreleased model based on Qwen 3.8 27B scored 77.8%, just above Sol 6.1 at 76.9%. The Figma agent uses its in-house model for precise edits and iterations, while relying on third-party models for tasks such as image generation and new design exploration. Since beta, typical tasks have become about 20% faster; the agent beats or matches its predecessor in 62% of expert comparisons, earns 33% higher user ratings, and sees 60% fewer undo actions. Figma also says its own model powers more than twice the share of work and average message costs have fallen, reflecting improvements across the broader agent system.