HOW WE'RE TRAINING A FIGMA-FLUENT EDITING MODEL

Summary from the TLDR Design newsletter

Specialized post-training lets Figma's first-party editing model compete with frontier models on precise, iterative edits to existing designs. A Llama 3.3 70B model trained on 250,000+ examples first became competitive with top models, but imitation learning limited generalization, so Figma now invests in reinforcement learning. Since beta, typical tasks run about 20% faster, users rate results 33% higher, and they undo the agent's work 60% less often.