Figma Trained Llama on 250,000 Examples as Canva Ships Brush

Figma said on 29 October that it trained “one of our earlier Llama 3.3 70B models” on “more than 250,000 examples” and got it “to be competitive on editing tasks with top models at the time”, in an article it posted on X about its own design-editing model.

Figma’s article went up at 19:51 UTC, 34 minutes after Canva introduced Brush, its own vision-language model for design edits, at 19:17 UTC, by our arithmetic from the two posts’ timestamps. Both companies describe a model trained for one job: changing designs that already exist.

Short explainer · Figma and Canva models

A vertical news Short narrated from this piece’s headline and opening line.

Business Pill · Model Routing

An evergreen Business Engineer explainer on model routing and inference cost. It explains the general idea only and does not discuss Figma or Canva.

What Figma Disclosed

Figma says that when it started, “frontier models couldn’t understand Figma syntax, and getting this to work would take a lot of system prompting that would be incredibly token-consumptive”. It turned instead to supervised fine-tuning, pairing a starting design with a synthetic edit request and a successful example of the change, written in Figma’s native file format.

Larger teacher models attempted the edits, and Figma says it checked their output with “deterministic validators and visual model judges”. Since then, the article says, Figma has “moved towards more capable base models including the newer Qwen models, Muse Spark Open, Inkling, and others” and is investing in reinforcement learning, because imitation learning alone came at “the cost of loss of generalization”.

The CEO’s Post

Figma co-founder and CEO Dylan Field shared the article on X late on 29 October, above, writing: “You would be shocked if I told you the things we made Llama 3.3 70B do.” His post had about 16,000 views when we read it on 30 October; Figma’s article had about 52,000.

How Figma Uses Its Model

Figma says it rolled out the Figma agent to all users “earlier this month” and that it “includes our first-party editing model”. The agent runs that model alongside “leading third-party models”: Figma says it uses the best third-party model for image generation, vectors and new design explorations, and its own model “for precise edits and iterations that our model handles well”.

Figma says it judges candidate models on how well they work in Figma and follow instructions, “but also the speed and cost of doing that work”. It concludes that “post-trained models can compete with frontier models on a specialized task like Figma editing”.

Canva Announced Brush 34 Minutes Earlier

Canva says Brush “started from a leading open-weights model”, post-trained on design and editing tasks, and is “up to 69% faster for general design editing tasks and 46% lower cost per edit task compared to Canva’s previous design agent”. Our Canva Brush piece covers those figures and their baseline.

The two companies disclose different things. Canva published speed and cost figures against its own previous agent; Figma published its training method and the size of one earlier training set, but no cost or speed figures in the article’s text. Our Figma business model page and our March analysis of Figma cover how the company makes money.

The Structural Read

As we read the two announcements, both companies describe pairing a model of their own for editing existing designs with models from other providers: Figma names third-party models, and Canva says it combines leading partner models with its own.

As we read Figma’s article, the stated reason was cost as well as quality: it says frontier models needed system prompting that would be incredibly token-consumptive.

By our arithmetic from the timestamps, Canva’s post went up at 19:17 UTC and Figma’s at 19:51 UTC on 29 October; neither company refers to the other.

Three Implications

FOR FIGMA USERS Figma says its agent, rolled out to all users earlier this month, includes its first-party editing model for precise edits and iterations.

FOR MODEL PROVIDERS Figma says it still uses the best third-party model available for image generation, vectors and new design explorations.

FOR INVESTORS Neither Figma’s article nor Field’s post gives a cost or speed figure; Canva gives its figures only against its own previous design agent.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework puts it this way: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.

As we read it, Figma’s article fits that pattern: it names open base models such as Llama 3.3 70B and the newer Qwen models, and adds its own training examples, validators and expert-written criteria. The framework was published in March 2026 and does not discuss Figma.

What Is Not Established

Figma does not say which base model its current production editing model uses, and the 250,000-example Llama 3.3 70B model is described as one of its earlier models. The article shows an internal evaluation chart, but this piece does not reproduce scores from it. Canva’s figures are its own, measured against its previous design agent. The 34-minute gap is our arithmetic from the posts’ timestamps; neither company refers to the other.

The Bottom Line

Figma says one of its earlier Llama 3.3 70B models, trained on more than 250,000 examples, was competitive on editing tasks with top models at the time, and that its agent now runs a first-party editing model next to third-party ones. By our arithmetic, its article went up 34 minutes after Canva introduced Brush, which Canva says it post-trained from an open-weights base.