In an official blog post on October 29, 2026, Figma described how it trains a dedicated canvas editing model for the Figma agent. The starting point is deliberately narrow: not generating a design from scratch, but helping designers change what is already on the canvas — adjusting spacing, changing a color, reworking a layout after feedback. These are small edits repeated all day, and they are hard to judge with a single standard.
More than 250,000 synthetic examples to learn the edits first
Figma began with supervised fine-tuning. Each training example paired a starting design with a synthetically generated edit request and one successful change, represented in Figma's native file format so the model's output could be applied back to the canvas. Larger teacher models attempted the edits first, and deterministic validators and visual model judges checked whether the output was valid and fulfilled the request; the pipeline could revise an edit using evaluator feedback. Figma says an early model based on Llama 3.3 70B, trained on more than 250,000 examples, became competitive with the top models of the time on editing tasks.
Imitation reached its limit, so reinforcement learning adds generalization
Imitation alone has a cost: the model tends to reproduce designs it saw in training, generalization weakens, and covering every kind of edit takes a large number of examples. Figma has therefore added reinforcement learning to its latest training recipe, focusing on RL environments closer to the design domain, more diverse examples, better-shaped reward models, and expert-written criteria so that rewards reflect what designers actually care about. The base models have been updated too, and Figma mentions newer Qwen models among them.
Evaluation is scored dimension by dimension, and the product numbers come first
Figma does not reduce design quality to a single score. It evaluates alignment, contrast, request fulfillment, and correct component use separately, combining structural checks with visual judges and calibrating the results with expert designers. An internal evaluation chart published by Figma shows its latest unreleased editing model at 77.8%, slightly above another model at 76.9%. On the product side, Figma says the agent has rolled out to all users, typical tasks became about 20% faster from beta to general availability, the new agent beat or matched the previous production agent in 62% of head-to-head tests with expert designers, user ratings rose 33%, and users undid the agent's work 60% less often.
