← Back to feed

Figma has published how it trains the model behind its design agent, and the headline is a split architecture: its own post-trained model handles precise edits, while third-party frontier models take image generation, vectors and brand-new design explorations. The newest unreleased edits model is post-trained on Qwen 3.8 27B.

The agent numbers

Between beta and general availability, Figma reports:

• Typical tasks 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.
• User ratings 33% higher.
• Users undo the agent’s work 60% less often.
• The share of work powered by Figma’s own post-trained model has more than doubled, and average cost per message has fallen.

Figma attributes those gains to the models, tools and harness together, not the editing model alone. No absolute usage share or cost figures were given.

How the model is built

The first pipeline was supervised fine-tuning. Each example paired a starting design with a synthetic edit request and a successful result, with designs represented in Figma’s native file format so outputs could be applied directly back to the canvas. Larger teacher models attempted the edits, and candidates were filtered by deterministic validators and visual model judges. The pipeline could revise a failed edit using evaluator feedback.

Figma says an early Llama 3.3 70B model trained on more than 250,000 examples was competitive with top models at the time on editing tasks. Frontier models then could not read Figma’s syntax without long, token-heavy system prompts, which is why Figma fine-tuned rather than prompted.

The team has since moved to newer base models, including newer Qwen models, Muse Spark Open and Inkling. Figma names two limits of SFT alone: imitation learning produces designs close to the training set at the cost of generalisation, and it needs many examples across the facets of a task. The current recipe adds reinforcement learning, with work on design-specific RL environments, more diverse training examples, reward models, and expert-written criteria.

How edits are scored

Figma evaluates structure and appearance separately, combining structural checks with visual judges calibrated by expert designers. Its examples of failures are overlapping icons and text, clipped arrows and the wrong component used for a button that looks right. The company published a comparison chart of the unreleased Qwen 3.8 27B model against third-party models on its internal editing evals but describes it as one input among many; the underlying scores are internal and unaudited.

The point

A 27B open-weights model, post-trained on a narrow task with a proprietary file format and a proprietary reward stack, is being routed the repetitive part of a product used by designers daily, while frontier APIs keep the open-ended work. Figma says the work shows post-trained models can compete with frontier models on a specialised task. The claim rests on internal evals and has not been independently tested.