Figma has published research on the post-trained model that powers the editing side of its Figma agent. The work centres on a narrow, frequent class of tasks — adjusting spacing, changing a color, reworking a layout after feedback — where changes are small, iterative and generally non-verifiable, and each edit builds on earlier decisions. When building the agent, the team focused on helping designers edit work already on their canvas, with speed treated as a first-class requirement: an edit that takes too long interrupts the workflow it is meant to support.
Why prompting gave way to fine-tuning
At the outset, frontier models could not understand Figma syntax, and making them work through system prompting would have been extremely token-consumptive. The team instead invested in supervised fine-tuning (SFT). Each training example paired a starting design with a generated synthetic edit request and a successful version of the change, and designs were represented in Figma’s native file format so the model would learn to produce edits applicable back to the canvas. Where tasks needed design system assets, synthetic prompts covered changes to themes and components.
Larger teacher models then attempted those edits, with candidates checked by deterministic validators and visual model judges — including checks for valid output and request fulfilment — and the generation pipeline could self-heal and revise an edit using evaluator feedback. The approach carried further than expected: an earlier Llama 3.3 70B model trained on more than 250,000 examples became competitive on editing tasks with top models at the time.
The team has since moved to more capable base models, including newer Qwen models, Muse Spark Open and Inkling. SFT alone, however, hit limits for this objective. Imitation learning pushes the model toward designs resembling its training data at the cost of generalisation, and it is less sample efficient, requiring large numbers of examples to cover the task’s facets. The latest training recipe therefore also incorporates Reinforcement Learning (RL), with effort directed at better RL environments for the design domain, greater example diversity, reward model shaping and expanded expert-written criteria so rewards reflect what designers actually care about.
Scoring an edit on separate axes
Candidate models go through internal evals covering Figma fluency, instruction following, usefulness of changes to existing designs, speed and cost. Quality is treated as several measurable dimensions rather than one: a design can have good alignment and poor contrast, an edit can look polished while missing the request, and a button can look right while using the wrong component. Structural checks on the design are combined with visual judges, and expert designers calibrate and review the results — a process meant to show where a model succeeds and what it still needs to learn. Scores from top models on these internal editing evals are one input among many in deciding which models the product considers.
Inside the shipped agent
The Figma agent, rolled out to all users earlier this month, includes the first-party editing model. It runs both that specialised model and leading third-party models, splitting work by task: image generation, vectors or brand-new design explorations go to the best available third-party model, while precise edits and iterations the in-house model handles well use Figma’s own. Canvas and design-system context, plus tools that can carry out the requested changes, matter alongside the model; post-training improves the model’s part of that work.
Reported results from beta to 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.
- Users undo the agent’s work 60% less often.
The share of work powered by Figma’s own post-trained design model has more than doubled, and average cost per message has fallen. Those gains span the models, tools and harness; the editing model is one component, and it gives the team a lever for the tasks designers repeat most.
What the results suggest
The work indicates that post-trained models can compete with frontier models on a specialised task like Figma editing, and that starting from a focused problem turns progress into product updates — leaving designers less time managing individual edits. Figma says it trains in a way that respects customer privacy and preferences, and it is hiring for design and post-training roles in non-verifiable domains.



