Excited to share more about the work behind Figma's first-party model: a specialized model for design editing that competes with top models. When we set out to build an agent, we wanted a model that understands design concepts and Figma's tools to help you quickly make edits on the canvas. We started with supervised fine-tuning (SFT) using Figma's native file format, pairing designs with synthetic edit requests. Deterministic validators, visual judges, and a self-healing pipeline helped evaluate and refine our outputs. We've since added reinforcement learning (RL) to our recipe to push for more improvements. Design is subjective, but we can still evaluate structural correctness, visual alignment, and request fulfillment using judges calibrated with expert designers. That includes checking whether an edit preserves everything it should. AI models often create unintended changes like dropped text or mismatched chart labels that we need to catch. Proud of the work of our incredible AI research team. Read more about the model here, and if this work sounds interesting, come join the team!