Figma's in-house editing model now beats top frontier models at design edits
We're Training a Figma-Fluent Editing Model

Figma built a first-party editing model to handle the small, iterative, non-verifiable tweaks designers make all day: spacing, colors, layout fixes. Starting with supervised fine-tuning on Figma's native file format, the team got a Llama 3.3 70B model trained on over 250,000 examples to compete with top models. Now they're adding reinforcement learning, better environments, and expert-shaped rewards. The agent is about 20% faster, users undo its work 60% less often, and ratings are up 33%.
We were even able to get one of our earlier Llama 3.3 70B models, trained on more than 250,000 examples, to be competitive on editing tasks with top models at the time!
- jjcm
Ex-Figma.
For context, this blog was released the same day Canva made their announcement for Brush, which looks to be further along in this area: https://www.canva.com/newsroom/news/canva-brush/
I'm glad Figma is pushing in this area. I do think that Figma has one of the best UI datasets out there in the world, and for the most part they've been sitting on it letting it get unused. Excited to see where they take this, but I do think they should be looking at image models in addition to LLM-based edit models (I'm biased as I left figma to start a image-mode-based UI design tool) as I find they're more capable in this area.
- miguel-muniz
I've used various combinations of Codex, Claude Code, GitHub Copilot, Figma's MCP, Figma Agent, Pen.dev, and Paper.design for product design. I have previous experience with front-end dev so I'm more technical than the average product designer.
In my experience, Figma Agent and Figma's MCP leave something to be desired. I'm sure Figma Agent is useful for people who haven't used Claude Code or similar before, but it really requires most of your product's context to exist in Figma for it to do well.
As soon as your product's context exists out of Figma, then it starts to drift and hallucinate its own styles. Using the MCP with an AI that has access to your code provides better results, but even it will have trouble remembering what page I'm asking it to edit.
I know they're trying to address this by adding features such as guidelines and specs, along with linking your Figma files to GitHub repos, but my complaint with these features is a lot of the companies I have worked for do not use the very specific tech stack that Figma supports with these features. If it's anything like code connect, it may take years to get more broad support. Every company I have worked for is a C# Microsoft/Azure house.
This does not help with my broader frustration with Figma which is there is a lack of focus/vision. They're trying to compete on all fronts by developing all these different products/features and it doesn't really feel like they have a cohesive story for what the future or product desi […]
- amiantos
I might be wrong or mistaken, but it seems like a big disadvantage for Figma that in order to use an agent with Figma, it has to be a specially trained Figma-aware agent and not just whatever agent I personally want to use.
Feels to me like a big issue that a smart competitor could take advantage of. I want to use my agent, I don't want to pay Figma to use their agent.
I'm also a bit salty about Figma ratcheting down hard on API calls, which creates customer service issues for any product that integrates with the Figma API. So maybe I am biased, but based on comments in here it also seems like the Figma MCP isn't very good either. Is Figma's long-term strategy to go full walled garden?
- taitems
Anything would be better than their current embarrassing AI roll out. The only way you can get something halfway decent is to put Claude Code into Fable or some equally high intent window and make an animal sacrifice.
Then the whole thing is effectively useless because of server-side font licensing. Your expensive brand font cannot be used by the MCP so you then need an additional layer of client-side Figma plugins to replace the placeholder font with your real font.
- ralusek
1.) They should have been doing this 3 years ago
2.) You can use any agent with Figma
3.) You can use any agent with not-Figma
I used to really respect Figma's engineering team and have been blown away by the degree to which they've dropped the ball with AI.
All they need is to expose their existing API very well over MCP and give a "screenshot" feature agents can utilize and the agents are good to go.