Figma explains how it trains its own design-editing model for the Figma agent

Figma has published a post by Josh Clemm, its Vice President of Engineering, on how it trains the model behind the Figma agent, the design agent that works directly on the canvas. Figma's argument is that specialized post-training can compete with frontier models, and the post shares the research behind applying that to design editing.

The starting point is the kind of work designers do all day: adjust the spacing, change a color, rework a layout after feedback. These edits are small, iterative and often non-verifiable, and each builds on earlier decisions. Figma wanted the agent to be better at design and faster at these edits, because a slow edit interrupts the workflow it is meant to help.

When the work started, frontier models could not understand Figma syntax, and making it work through prompting would have needed a lot of system prompting that would be incredibly token-consumptive. So Figma invested in supervised fine-tuning (SFT). Each training example paired a starting design with a generated synthetic edit request and a successful example of the change. Designs were represented in Figma's native file format, so the model learned to produce changes that could be applied back to the canvas. For tasks that needed design system assets, Figma generated synthetic prompts covering changes to themes and components.

To get the successful examples, Figma used larger teacher models to attempt the edits. Candidates were checked with deterministic validators and visual model judges, including checks for valid output and for whether the request was fulfilled. The generation pipeline could self-heal and revise an edit using evaluator feedback. The post says this approach got the team surprisingly far: one earlier Llama 3.3 70B model, trained on more than 250,000 examples, was competitive on editing tasks with top models at the time.

Since then Figma has moved to more capable base models, including the newer Qwen models, Muse Spark Open, Inkling and others. It also hit limits with SFT alone. Imitation learning teaches the model to produce designs very similar to what it saw in training, at the cost of generalization, and it is less sample efficient, so training needs many examples covering the facets of the task. Figma is therefore also investing in reinforcement learning (RL) as part of its latest recipe. The stated focus is better RL environments for the design domain, more diverse training examples, shaping the reward models, and expanding expert-written criteria so rewards reflect what designers care about.

On evaluation, Figma runs candidate models through internal evals that measure how well a model works in Figma, follows instructions and makes useful changes to existing designs, plus the speed and cost of doing so. Quality is not treated as one unmeasurable thing: a design can have good alignment and poor contrast, an edit can look polished and still miss the request, and a button can look right while using the wrong component. Figma evaluates these dimensions separately, combining structural checks with visual judges, and expert designers calibrate and review the results. The post shows how several top models score on the internal evals, but calls that only one input among many used to decide which models to consider for the product.

On results, the post says Figma rolled out the agent to all users earlier this month. Between beta and general availability, typical tasks are about 20% faster. In head-to-head testing with expert designers, the new agent beats or matches the previous production agent 62% of the time. User ratings are 33% higher, and users undo the agent's work 60% less often. Figma has more than doubled the share of work powered by its own post-trained design model and has reduced the average cost per message. The post attributes these gains to improvements across the agent, including the models, tools and harness, and describes the editing model as one part of that progress.

The post closes by saying the work has shown that post-trained models can compete with frontier models on a specialized task like Figma editing. It adds that Figma is hiring for people interested in design and post-training models for non-verifiable domains, and states that it trains in a way that respects customer privacy and preferences.

Key facts

Figma trained its design-editing model with supervised fine-tuning on synthetic examples: a starting design, a generated edit request and a successful edit, in Figma's native file format.
Larger teacher models produced the edits, which were checked by deterministic validators and visual model judges; an earlier Llama 3.3 70B model trained on more than 250,000 examples was competitive with top models of the time.
Figma has moved to newer base models (Qwen models, Muse Spark Open, Inkling and others) and is adding reinforcement learning because SFT alone loses generalization and is less sample efficient.
Reported early results for the Figma agent: typical tasks about 20% faster between beta and general availability, new agent beats or matches the previous one 62% of the time in tests with expert designers, ratings 33% higher, undos 60% less frequent.
Figma says it has more than doubled the share of work handled by its own post-trained model and lowered average cost per message, crediting improvements across models, tools and harness.

Why it matters

Figma is a design tool company arguing that a model post-trained on a narrow task can compete with frontier models on that task. Its case rests on a concrete recipe: synthetic edit data in Figma's native format, validation by teacher models and judges, then reinforcement learning with design-specific rewards. The post says that when the work started, frontier models could not understand Figma syntax and prompting around that would have been very token-heavy, which is why Figma chose to fine-tune. It also says the share of work powered by its own model has more than doubled and average cost per message has fallen.

Who it affects

Designers using the Figma agent are the direct audience: the agent is now available to all users, and the post says its changes target the repetitive edits they make most often. Teams that build domain-specific models for tasks without a clear right answer (the post calls these non-verifiable domains) can read it as a description of one company's approach. The post also names teams at Uber, Granola and Atlassian as users of the generally available agent.

How to use it

This is a research and engineering write-up, not a product launch for developers. Designers can use the Figma agent directly on the canvas, as it is generally available to all users. For practitioners, the transferable ideas in the post are: generate synthetic edit requests paired with successful edits, validate candidates with deterministic checks and visual judges, evaluate quality dimensions such as alignment, contrast, request fulfillment and correct component use separately, and have expert designers calibrate the judges. The post does not describe any release of weights, datasets or tooling.

How solid is it

It is a first-party post by a Figma executive, and every figure is Figma's own. The post calls the agent results early. It gives no baseline or measurement method for the 20% speed gain, the 33% rating increase or the 60% drop in undos. The 62% figure comes from head-to-head testing with expert designers against the previous production agent, and the post says the gains come from improvements across models, tools and harness, not the editing model alone. The score comparison of top models on Figma's internal evals is shown only as a chart, and Figma itself calls it just one input among many.

Risks and caveats

The reduction in cost per message is not quantified, and the absolute share of work handled by Figma's own model is not given. No RL results are reported; RL is described as an investment and focus, not as a measured outcome. The post does not name the teacher models or the visual judge models, does not say which base model powers the production editing model today, and does not state the size of the final model. The Llama 3.3 70B claim refers to an earlier model compared with top models at that time, not with current ones. The claim that post-trained models can compete with frontier models is Figma's conclusion about one specialized task, Figma editing.

“This work has demonstrated that post-trained models can compete with frontier models on a specialized task like Figma editing.”

— Josh Clemm, Vice President of Engineering, Figma