Figma has built its own design-editing model and put it inside the Figma agent, where it takes the small, repetitive changes that fill a designer’s day. In a post on the Figma blog, Josh Clemm describes a model post-trained on top of open base models, which the agent uses for precise edits and iterations. Bigger third-party models still get the jobs that start from nothing.
The split is the interesting design decision. For image generation, vectors, and brand-new design explorations, the agent calls the best outside model available. For things like adjusting spacing, changing a color, or reworking a layout after a colleague leaves feedback, it uses Figma’s own. Those edits build on earlier decisions in the file, and an agent that takes too long to make them gets in the way of the work it was meant to speed up.
The training began with supervised fine-tuning. Every training example combined a starting design, an invented edit request, and a version of the change that worked. Designs were encoded in the format Figma itself saves files in, so whatever the model produced could be applied straight back to the canvas. Larger teacher models attempted the edits, and automated validators plus visual judges filtered the results. Figma says an early Llama 3.3 70B model, fed over 250,000 of these examples, became competitive with the best models then available at editing.
Imitation has a ceiling, though. Clemm writes that a model trained only to copy examples drifts toward the designs it has seen and generalizes worse, and that it needs a lot of examples to learn each facet of the task. Figma has since moved to newer bases, including Qwen models, Muse Spark Open, and Inkling, and is adding reinforcement learning. That means building practice environments for design, widening the range of training examples, and shaping reward models around criteria written by expert designers.
Grading is the hard part, because a design can fail in several unrelated ways. A layout can be well aligned with poor contrast, an edit can be polished and still ignore the request, or a button can look correct and still be built from the wrong component. Figma scores those separately, mixing structural checks with visual judges and having expert designers calibrate the results. Its examples of typical failures are overlapping icons and text, clipped arrows, and overlapping text elements. Figma’s newest edits model, which is not yet released and was built on Qwen 3.8 27B, is compared against third-party models in a chart on the blog, though the post itself reports no scores in text.
None of the product numbers have outside verification. Figma says typical tasks run about 20% faster than they did in beta. Expert designers who compared the new agent with the version it replaced preferred it or called it equal in 62% of cases. Ratings from users are up 33%, users reverse the agent’s changes 60% less often, and the share of work handled by Figma’s own model has more than doubled. Average cost per message has also fallen. Figma credits all of it to the models, tools, and harness together, so the editing model’s own share of the gain is unknown.
The comparison that matters for anyone building an AI feature inside an existing product is the one Figma chose not to make. It does not claim its model beats frontier models at design in general, only that a small model trained hard on one repeated task can hold its own there, at lower cost and latency. A product team with a few high-volume actions and a way to grade them now has a reference for routing those actions to a cheaper, specialized model, and keeping the expensive one for the open-ended requests.
Reported by the Figma blog, in a post by Josh Clemm (no publication date given in the source).
