AI BriefingFigma's unreleased editing model, post-trained on Qwen 3.8 27B, beats top third-party model on internal evals
·2026.10.30 06:00
Key point
Figma shares how its latest unreleased editing model, post-trained on Qwen 3.8 27B, scores 77.8% on internal evals, surpassing the top third-party model, while the general availability agent sees 20% faster task completion.
Details
- Figma's latest unreleased editing model, post-trained on Qwen 3.8 27B, scores 77.8% on internal design editing evals, surpassing the top third-party model (Sol 6.1) at 76.9%.
- The model uses supervised fine-tuning (SFT) and reinforcement learning (RL) to handle iterative, non-verifiable design edits like spacing and color changes.
- Since the Figma agent's general availability, typical tasks are 20% faster, and the agent beats or matches the previous production version 62% of the time in expert designer tests.
- User satisfaction metrics show ratings are 33% higher and users undo the agent's work 60% less often compared to the previous version.
- Figma has more than doubled the share of work powered by its own post-trained model, reducing the average cost per message.
- The training pipeline uses synthetic edit requests, deterministic validators, and visual model judges to ensure valid output and request fulfillment, with expert designers calibrating the evaluation criteria.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
The Escape Room EconomyPerplexity Proposes JOLT: Joint On-Policy Learning and Teaching for Efficient RL Post-Training
Mushinseok Developer: The Limit of Parallel AI Agent Work Is Judgment, Not Coding
Cohere Reports Qwen3.8-Flash-Next Attempting Unauthorized Network Uploads in Sandbox