Canva ships Brush and Figma shows its edit model, as design tools post-train their own

Two design companies explained their in-house AI editing models within a day of each other, and the recipes look alike.

Canva Brush

Canva introduced Brush, a vision-language model for design built by Canva Research, its 160-person AI lab. Brush looks at an existing design, works out how it is put together, and edits it on request, from one line of copy to every slide in a deck. Canva says it started from “a leading open-weights model,” which it did not name, then post-trained it on real design and editing tasks. That included anonymised data from users who opted in, and reinforcement learning with a reward system that scores visual quality, layout and whether the edit did what was asked. Behind the scenes it runs an agent loop that can check and improve its own work.

Canva’s numbers against its previous design agent: up to 69% faster on general design editing and 46% lower cost per edit task, and in internal testing it matched or beat the old setup on most edits. Brush is live now for Canva Pro subscribers. To make sure it handles your edits, pick “Brush” in the Canva AI model menu. Brand Kit support, attachments as references and better handling of image-heavy layouts are listed as next. Canva pairs it with Craft, its own image model: Craft makes the pictures, Brush edits the design around them.

Figma’s edit model

On 29 October Figma’s Josh Clemm wrote up how Figma trains a “Figma-fluent” editing model. The team generated more than 250,000 synthetic edits in Figma’s native format and used supervised fine-tuning first; an early run on Llama 3.3 70B was already competitive. Later rounds tested Qwen, Muse Spark Open and Inkling bases and added reinforcement learning. A chart caption says the latest model, not yet released, is post-trained on Qwen 3.8 27B.

Figma reports that between beta and general availability its agent became about 20% faster on typical tasks, beat or matched the previous agent 62% of the time, earned 33% higher user ratings, and got undone 60% less often. Those are Figma’s own measurements, and they cover the agent as a whole, not the model alone.

Why this matters

Both companies are betting that a mid-sized open-weights model trained on their own format and their users’ edits beats calling a general frontier model for every change, on speed and cost at least. If you build editing features on a structured document format, that is the pattern to test: synthetic edits for SFT, then a reward model for the format’s rules. As with Microsoft’s edit-first MAI Image 2.7, the claims here are vendor-measured, so try your own files before switching.