---
id: "lib-2026-10-29-how-we-re-training-a-figma-fluent-editing-model"
type: "library"
title: "How we're training a Figma-fluent editing model"
lang: "EN"
org: "Figma"
source_type: "Practitioner post"
date: "2026-10-29"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-29-how-we-re-training-a-figma-fluent-editing-model/"
original_url: "https://www.figma.com/blog/training-a-figma-fluent-editing-model/"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Figma's engineering VP describes how the company post-trains its own design-editing model; a credible view of a product team's training recipe, but results come from internal evals and product metrics with no method or sample sizes."
summary_basis: "full-text"
original_minutes: "6"
tags: ["ai technology", "future of work", "enablement", "business transformation", "small specialised models", "post training", "design tools"]
---

# How we're training a Figma-fluent editing model

Tier B (Useful, limited): Figma's engineering VP describes how the company post-trains its own design-editing model; a credible view of a product team's training recipe, but results come from internal evals and product metrics with no method or sample sizes.

Summary based on the full text.

## In short

A Figma blog post (29 October 2026) by Josh Clemm, VP of Engineering, on the editing model inside the Figma agent. Frontier models did not understand Figma's file format without heavy, costly prompting, so Figma post-trained its own model for small, repeated edits (spacing, colour, layout changes) and routes other work, such as image generation and new explorations, to third-party models. The latest version, built on Qwen 3.8 27B, slightly edges a leading closed model on Figma's own editing eval.

## Key numbers

- 77.8% vs 76.9%: internal Figma editing eval: Figma's model post-trained on Qwen 3.8 27B versus Sol 6.1
- 250,000+: supervised fine-tuning examples used for an earlier Llama 3.3 70B editing model
- 62%: share of expert head-to-head tests where the new agent beats or matches the previous production agent
- 60%: fewer undos of the agent's work by users (ratings 33% higher, typical tasks about 20% faster)

## What it reports

- First recipe: supervised fine-tuning on synthetic triples (starting design, generated edit request, successful edit) in Figma's native format. Larger teacher models produced candidate edits, filtered by deterministic validators and visual model judges, with self-repair from evaluator feedback.
- An earlier Llama 3.3 70B model trained on more than 250,000 examples was competitive with top models of its time.
- Fine-tuning alone hit limits: weaker generalisation and poor sample efficiency. The current recipe adds reinforcement learning, with work on design environments, reward models and expert-written criteria.
- Quality is scored per dimension (structure, alignment, contrast, request fulfilment, correct components), combining structural checks, visual judges and expert designer calibration.
- Internal eval: Figma design edits v3 at 77.8% versus Sol 6.1 at 76.9%.
- Product results between beta and general availability: typical tasks about 20% faster, user ratings 33% higher, 60% fewer undos, new agent beats or matches the old one 62% of the time; share of work on Figma's own model more than doubled and cost per message fell.

## How they got it

Internal evals and product telemetry. The post shows a training-pipeline diagram and one eval chart; it does not publish eval size, task mix or the rating method.

## Limits

- All figures are internal and unaudited; the eval is Figma's own.
- The product gains are attributed by the authors to models, tools and harness together, not to the editing model alone.
- The eval gap over the closed model is under one point.

## Why it matters for companies

A software vendor with a narrow, high-frequency task and its own data found that a mid-sized open model, post-trained, matches closed frontier models on that task while costing less per message. The pattern for buyers and builders alike: keep frontier models for open-ended work, and consider specialised models for repetitive edits where speed and cost per call dominate. Per-dimension quality scoring is a reusable idea for any work that lacks a single correct answer.

Original: https://www.figma.com/blog/training-a-figma-fluent-editing-model/