--- title: How we’re training a Figma-fluent editing model | daily.dev description: Figma built a specialized post-trained model for editing designs on the canvas, used inside the Figma agent. The team started with supervised fine-tuning,... canonical: https://daily.dev/posts/how-we-re-training-a-figma-fluent-editing-model-nmb5gc1oh twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How we’re training a Figma-fluent editing model | daily.dev og:description: Figma built a specialized post-trained model for editing designs on the canvas, used inside the Figma agent. The team started with supervised fine-tuning,... og:url: https://daily.dev/posts/how-we-re-training-a-figma-fluent-editing-model-nmb5gc1oh og:image: https://api.daily.dev/og/posts/NMb5Gc1OH.png og:image:alt: How we’re training a Figma-fluent editing model og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # How we’re training a Figma-fluent editing model **[Figma](https://daily.dev/sources/figma)** · 6 min read · 2 upvotes · 1 comments ## Summary Figma built a specialized post-trained model for editing designs on the canvas, used inside the Figma agent. The team started with supervised fine-tuning, training an earlier Llama 3.3 70B model on more than 250,000 synthetic edit examples to be competitive with top frontier models. They've since moved to newer base models like Qwen and Muse Spark Open, and are investing in reinforcement learning because SFT alone has limits on generalization. Since the agent's general availability, typical tasks are about 20% faster, the new agent beats or matches the previous production agent 62% of the time in head-to-head testing with expert designers, and users undo the agent's work 60% less often. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How did Figma train its own AI model to edit designs instead of just using a frontier LLM? Figma used supervised fine-tuning, pairing starting designs with synthetic edit requests and successful example changes, represented in Figma's native file format. Larger teacher models attempted the edits, with deterministic validators and visual judges checking outputs, and a self-healing pipeline revising edits using evaluator feedback. An earlier Llama 3.3 70B model trained on over 250,000 examples became competitive with top frontier models on editing tasks. _Following how teams fine-tune models for narrow product tasks? daily.dev tracks posts like this one on AI post-training._ ### Why would reinforcement learning work better than supervised fine-tuning for an AI design editing model? Supervised fine-tuning alone causes imitation learning, where the model reproduces designs similar to training data but loses generalization, and it is less sample efficient, requiring many examples to cover task facets. Figma is instead investing in reinforcement learning, focusing on better RL environments for design, example diversity, reward model shaping, and expert-written criteria reflecting what designers actually value. _Developers weighing SFT versus RL for their own fine-tuning projects can follow this trade-off on daily.dev._ ### What measurable improvements came from Figma's new post-trained editing model in its design agent? Typical tasks became about 20% faster between the beta and general availability of the Figma agent. In head-to-head testing with expert designers, the new agent beat or matched the previous production agent 62% of the time, user ratings rose 33%, and users undid the agent's work 60% less often. _Anyone evaluating AI coding or design agents can track benchmark results like these on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@kartiknvj** · 0 upvotes > Beats or matches 62% of the time" puts wins and ties in one bucket, so I can't tell how much of it is ties. When I compare two agent versions I keep ties as their own count. Do the expert designers also judge head-to-head per dimension, say alignment and contrast on their own? ## Similar posts on daily.dev - [3 Ways Product Designers Use the Figma agent](https://daily.dev/posts/3-ways-product-designers-use-the-figma-agent-dsl76ad4t) · Figma · 2 upvotes · 0 comments - [Workflow lab: Staying in the flow with Figma’s agent](https://daily.dev/posts/workflow-lab-staying-in-the-flow-with-figma-s-agent-5z8zyezkg) · Figma · 1 upvotes · 0 comments - [How We Developed Zeta2 — Zed's Blog](https://daily.dev/posts/how-we-developed-zeta2-zed-s-blog-p5pf4r1qy) · Zed · 89 upvotes · 4 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#llm](https://daily.dev/tags/llm), [#deep-learning](https://daily.dev/tags/deep-learning), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning), [#figma](https://daily.dev/tags/figma) [View this post on daily.dev](https://daily.dev/posts/how-we-re-training-a-figma-fluent-editing-model-nmb5gc1oh) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"How we’re training a Figma-fluent editing model","url":"https://daily.dev/posts/how-we-re-training-a-figma-fluent-editing-model-nmb5gc1oh","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-we-re-training-a-figma-fluent-editing-model-nmb5gc1oh"},"datePublished":"2026-10-29T19:17:50.046Z","dateModified":"2026-10-29T19:18:19.092Z","description":"Figma built a specialized post-trained model for editing designs on the canvas, used inside the Figma agent. The team started with supervised fine-tuning,...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/0426a159f105ef80002df980f0735955?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/0426a159f105ef80002df980f0735955?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Figma","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Figma","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/d81fd85ddaea4d25a658694de448118f","url":"https://daily.dev/sources/figma"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/how-we-re-training-a-figma-fluent-editing-model-nmb5gc1oh","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"ai-agents,llm,deep-learning,reinforcement-learning,figma","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Figma","item":"https://daily.dev/sources/figma"},{"@type":"ListItem","position":3,"name":"How we’re training a Figma-fluent editing model"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/how-we-re-training-a-figma-fluent-editing-model-nmb5gc1oh","comment":[{"@type":"Comment","text":"Beats or matches 62% of the time\u0026quot; puts wins and ties in one bucket, so I can’t tell how much of it is ties. When I compare two agent versions I keep ties as their own count. Do the expert designers also judge head-to-head per dimension, say alignment and contrast on their own?","datePublished":"2026-10-30T13:34:47.825Z","url":"https://daily.dev/posts/NMb5Gc1OH#c-OmZIUWksQ","author":{"@type":"Person","name":"kartik-nvjk","url":"https://daily.dev/kartiknvj","image":"https://media.daily.dev/image/upload/s--3gGgsVCw--/f_auto/v1781456774/avatars/avatar_TvTVeiMdkRCqWUDullFmy?_a=BAMAMiWQ0"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/how-we-re-training-a-figma-fluent-editing-model-nmb5gc1oh#faq","mainEntity":[{"@type":"Question","name":"How did Figma train its own AI model to edit designs instead of just using a frontier LLM?","acceptedAnswer":{"@type":"Answer","text":"Figma used supervised fine-tuning, pairing starting designs with synthetic edit requests and successful example changes, represented in Figma's native file format. Larger teacher models attempted the edits, with deterministic validators and visual judges checking outputs, and a self-healing pipeline revising edits using evaluator feedback. An earlier Llama 3.3 70B model trained on over 250,000 examples became competitive with top frontier models on editing tasks. Following how teams fine-tune models for narrow product tasks? daily.dev tracks posts like this one on AI post-training."}},{"@type":"Question","name":"Why would reinforcement learning work better than supervised fine-tuning for an AI design editing model?","acceptedAnswer":{"@type":"Answer","text":"Supervised fine-tuning alone causes imitation learning, where the model reproduces designs similar to training data but loses generalization, and it is less sample efficient, requiring many examples to cover task facets. Figma is instead investing in reinforcement learning, focusing on better RL environments for design, example diversity, reward model shaping, and expert-written criteria reflecting what designers actually value. Developers weighing SFT versus RL for their own fine-tuning projects can follow this trade-off on daily.dev."}},{"@type":"Question","name":"What measurable improvements came from Figma's new post-trained editing model in its design agent?","acceptedAnswer":{"@type":"Answer","text":"Typical tasks became about 20% faster between the beta and general availability of the Figma agent. In head-to-head testing with expert designers, the new agent beat or matched the previous production agent 62% of the time, user ratings rose 33%, and users undid the agent's work 60% less often. Anyone evaluating AI coding or design agents can track benchmark results like these on daily.dev."}}]} ```