AI AI Toolkit
AI Newstip

Figma 讲解如何训练专用画布编辑模型

Figma AI / Agent / Make2026-10-29T21:00:00.000Z

Key Highlights

Figma engineering VP Josh Clemm explains how the team trained a model that speaks Figma. The core thesis is that specialized post-training can rival frontier models, at least on the narrow task of design editing. The Figma agent, generally available since early this month, ships this first-party editing model, and the write-up is unusually candid about the methodology behind it.

What Happened

Figma's starting point was a focused task: helping designers modify work already on the canvas, adjust spacing, change color, rework layout after feedback. These everyday edits are scattered, iterative, and often non-verifiable. The team found early frontier models simply did not understand Figma syntax, and propping them up with system prompts was both token-expensive and ineffective, so they turned to supervised fine-tuning rather than fighting the base model's ignorance.

Technical Details

Training examples paired a starting design with a synthetic edit request and a successful change, representing designs in Figma's native file format so the model learns to emit changes applicable back to the canvas. For design-system assets, synthetic prompts covered theme and component changes. Larger teacher models then attempted edits, gated by deterministic validators and visual-judge checks, valid output and request fulfillment, and the pipeline could self-heal and revise using evaluator feedback. An early Llama 3.3 70B trained on 250k-plus examples could already compete with top models of the time. Recently the team moved to stronger bases like Qwen 3.8 27B and added reinforcement learning to fix SFT's weak generalization, focusing on design-domain RL environments and reward models.

Versus Competitors

On internal Figma editing eval v3, Figma's post-trained model, on Qwen 3.8 27B, scored 77.8% versus Sol 6.1's 76.9%. From beta to GA, typical tasks were about 20% faster, beat or matched the prior agent 62% of the time with expert designers, user ratings were 33% higher, and undo rate 60% lower. Figma's strategy is clear, third-party models for image generation and vector exploration, in-house model for precise iterative editing, a split that optimizes both cost and control.

Industry Impact and Use Cases

This article is instructive for teams building vertical AI products. You need not blindly chase the largest frontier model. For high-frequency, verifiable scenarios, dedicated post-training is often cheaper, faster, and more controllable. Figma has more than doubled the share of work handled by its in-house model and lowered average cost per message, proving that small model plus good data plus post-training can beat general-purpose giants in a specialized domain.

Practical Notes

The reusable lesson is to scope the problem tightly before reaching for the biggest model. If a task is repetitive and has a checkable output, a fine-tuned smaller model will usually beat a frontier model on cost and latency while matching quality. Reserve the giant models for the genuinely open-ended parts, and let validators and judges do the hard verification work.

What to Watch

Watch whether Figma open-sources or productizes its post-training recipe. If the design-domain RL environment becomes a template, every vertical SaaS could clone this playbook for its own narrow, verifiable workflows and undercut the general-purpose labs on price.

Takeaway

Figma's editing model is proof that in a well-scoped domain, disciplined post-training beats raw scale, and it is a blueprint for vertical AI done right that other product companies should study closely.

Background

Figma's result is a data point in a larger debate about whether smaller specialized models can beat frontier models on their own turf, and the answer here is yes but only because the task is narrow, verifiable, and high-frequency enough to make the training data pay for itself. That specificity is the lesson most teams miss when they reach for the biggest model by default.

Where to Try

If you build a vertical product, identify your own narrow, verifiable loop and test whether a fine-tuned smaller model matches your frontier baseline on cost and latency. Figma's eval methodology, validators plus visual judges, is a template worth copying, because without a checkable signal you cannot tell whether your smaller model is actually as good.

Closing Thoughts

The strategic implication is that vertical SaaS companies may be able to detach from the frontier-lab price treadmill by owning a small model tuned to their domain, and that could reshape who captures the value in applied AI.