In a post on October 29, Figma disclosed a custom design-editing model. It is built on Alibaba's open-source Qwen 3.8 27B, refined with supervised fine-tuning and reinforcement learning, and tasked with making precise changes to existing designs. The model is already wired into Figma's AI agent, which is now generally available to all users.
The post gives the model no public name; the benchmark chart labels it "Figma design edits v3." On Figma's internal editing benchmark it scores 77.8%, against 76.9% for the third-party model Sol 6.1.
"We've increasingly seen that specialized post-training can compete with frontier models."
Two steps: imitate first, then reinforce
The first version used Llama 3.3 70B as its base. Each training example paired an original design file, a synthetic edit request and a successful edit result, with designs represented in Figma's own file format. Design-system assets such as themes and components also got synthetic prompts.
In the data pipeline, a larger teacher model first attempts the edit, and the result is then checked by deterministic validators and reviewed by a vision model. The pipeline can self-correct based on benchmark feedback. Early versions used more than 250,000 samples and could already match the leading models of the time on editing tasks.
The weakness of imitation learning soon showed: the model could only produce designs resembling its training samples, generalized poorly and was inefficient with data. The team then turned to reinforcement learning, focusing on design-specific training environments, sample diversity, reward models and scoring rubrics written by design experts. While moving the base model to Qwen 3.8 27B, they also evaluated newer Qwen models, Muse Spark Open and Inkling.
The numbers after launch
Figma shared several changes between the beta and the general release:
- Typical tasks run about 20% faster;
- In blind tests against professional designers, the new agent beat or matched the old one in 62% of comparisons;
- User ratings are 33% higher, and users undo agent edits 60% less often;
- The share of work handled by the in-house model has more than doubled, and the average cost per message has fallen.
The division of labor is spelled out too: precise edits and multi-turn iteration go to the in-house model, while generating images and vector graphics and exploring new designs from scratch still call the best-performing third-party models.
Others on the same road
More and more companies in the past two years have taken an open base model and post-trained it for one industry. Cursor trained its own models for code completion and rewriting, and Airbnb CEO Brian Chesky said publicly last year that the company's customer-service agent relies heavily on Qwen. Figma's thinking is similar: general-purpose models are good at creating from nothing, but making small, precise changes to someone else's design without breaking the design system takes a lot of in-domain data.
For readers following China's AI ecosystem, the interesting part is the base model. A US design-software company building its core editing capability on Qwen suggests that Chinese open-source models are now a viable option in commercial products overseas. Figma did not disclose how much costs fell, where the 27B model is deployed, or which cloud provider handles inference.
Sources: CocoLoop, Figma official blog; the official blog was used to verify the base model, training sample size and internal benchmark scores, and launch metrics reflect Figma's self-reported figures.