Figma's Editing Model Scores 77.8% Against 76.9% for Sol 6.1, and Only on Figma's Own Test
Software / news
Figma's Editing Model Scores 77.8% Against 76.9% for Sol 6.1, and Only on Figma's Own Test
The October 29 post names Qwen 3.8 27B as the base of an unreleased model, but explains neither the rival nor the test.

Figma said on October 29 that its in-house design-editing model scored 77.8% on an internal evaluation, 0.9 points ahead of a model it labels "Sol 6.1" at 76.9%. The post does not say what Sol 6.1 is, and the evaluation has not been released.
The post, "How we're training a Figma-fluent editing model", lists Josh Clemm, Figma's Vice President of Engineering, in its author section. His bio says he oversees Figma's AI platform and context layer. A chart caption in the post says the newest model is post-trained on Qwen 3.8 27B.

What the 77.8% compares
The only head-to-head figure in the post is a bar chart. Its alt text reads "Figma design edits v3 at 77.8% with Sol 6.1 at 76.9%". The 76.9% appears nowhere in the running text.
- Figma design edits v377.8 %
- Sol 6.176.9 %
Source: Figma blog post, October 29, 2026 (chart alt text), accessed 2026-11-04
Figma scores design structure and visual quality separately, combining structural checks with visual model judges. Expert designers calibrate and review the results. The post calls the evaluation "just one input among many."
A 0.9-point gap on a private test, with no sample size, is not a finding anyone outside Figma can check.
Product figures, all from Figma
The post also reports results from the shipped agent. Each is Figma's own measurement, and none comes with a sample size or a method.
| Measure | Figma's figure |
|---|---|
| Beats or matches the previous production agent, expert designers judging | 62% of the time |
| User ratings | 33% higher |
| Users undoing the agent's work | 60% less often |
| Typical task speed, beta to general availability | about 20% faster |
Figma also said the share of work handled by its own model has more than doubled and that the average cost per message has dropped. It gave no starting share and no cost.
Which base model is in production
The post names more than one base. It says an earlier Llama 3.3 70B model, trained on more than 250,000 examples, was "competitive on editing tasks with top models at the time". It then says Figma has "moved towards more capable base models including the newer Qwen models, Muse Spark Open, Inkling, and others."
The Qwen 3.8 27B tag applies to what the post calls the "latest unreleased Figma edits model". The post does not say which base the production model uses today.
The base model's Hugging Face card lists the licence as apache-2.0, a native context of 262,144 tokens and 8,014,877 downloads in the last month. The card metadata lists 28B parameters against the 27B in the name. The Terminal covered those downloads and Qwen's self-reported benchmarks earlier. Apache-2.0 on the base says nothing about the terms of Figma's derivative, whose weights are unreleased.
How the training changed
Figma's first approach was supervised fine-tuning. Each example paired a starting design, stored in Figma's native file format, with a synthetic edit request and a successful version of the change. Larger teacher models attempted the edits, and deterministic validators and visual judges checked them.
The post says imitation alone risked poor generalisation and was less sample efficient, so Figma now uses reinforcement learning with design-domain environments, reward models and expert-written criteria. Qwen models also feature in the dispute over MIT's weight-space programming claims.
What Figma has not said
Figma's built-in agent launched on May 20, 2026, according to The Next Web, which reported the company's claim that its models were "fine-tuned specifically for design work" and did not name them. The October 29 post is the first to name a base, and only for the unreleased model.
If you build on Figma's agent, nothing in the post lets you reproduce its numbers or choose the model. The post gives no release date for the editing model itself.
Sources
More in Software
- 01Levien's Next-100 Rust GUIs Post Counts About 20 New Toolkits and Bets Against egui's Immediate ModeThe Linebender author credits AI rewrites for the surge; egui's own README concedes the trade-off he is pressing on.
- 02C++26 Is Reopened: The 13–9–12 Ballot, the Enforced Label That Answers It, and What Is Still OpenSutter's Búzios report says the committee kept contracts, added one paper, and will send C++26 to a second ballot with no published date.
- 03OpenAI Says Codex Cannot Run in Docker for iOS Builds, and Docker Launches It With Codex's Sandbox OffBolin's November 18 post lists Seatbelt, Bubblewrap and MXC as Codex's three sandboxes; Docker's page defaults to --dangerously-bypass-approvals-and-sandbox.
- 04CLI-Anything Has 62,119 Stars and a README With Three Different Test TotalsThe HKUDS project turns desktop software into agent-callable CLIs; its own per-harness lines sum to 158 more tests than it reports.