The model achieved a 0.267 mean reward using KiCad.
🐨⚡️ @trycua had early access from @GoogleDeepMind to Gemini 3.5 Flash's native Computer Use! On Cua-Bench it posted the highest mean reward of any frontier model they tested -- at Flash speed and cost.
7/ Here's a demo our team recorded of Gemini 3.5 Flash driving a real KiCad task, via the @antigravity (agy) CLI and Cua Driver.
Users are excited about Gemini 3.5 Flash topping Cua-Bench with native computer use because of its speed, useful reasoning insights, and strong potential for automating tasks.
No Digg Deeper questions have been answered for this story yet.
7/ Here's a demo our team recorded of Gemini 3.5 Flash driving a real KiCad task, via the @antigravity (agy) CLI and Cua Driver.
8/ Blog: https://cua.ai/blog/evaluating-gemini-3.5-flash-on-computer-use
Try it: http://cua.ai/docs/cua-driver Benchmark: http://cua.ai/cuabench Repo: http://github.com/trycua/cua
@trycua @GoogleDeepMind amazing!! fast cua is so needed everywhere
1/ We had early access from @GoogleDeepMind to Gemini 3.5 Flash's native Computer Use. On Cua-Bench it posted the highest mean reward of any frontier model we tested - 0.267, on KiCad tasks no model fully solves. At Flash speed and cost.
@_philschmid @trycua @GoogleDeepMind Gemini 2.5 was my very first computer use model, and it changed my life! I started automating everything I could.
I have no doubt 3.5 will be exceptional
4/ The strongest signal was execution quality. Gemini 3.5 Flash grounded pixel-accurately on zoomed-in CAD targets, placed and edited parts, and kept working inside dense KiCad UI state.
@trycua @LLMJunky @GoogleDeepMind Thank you for testing 🤝
2/ The evaluation is Cua-Bench's KiCad EDA suite: 25 expert-authored professional CAD tasks, matched at a 200-step budget. The produced netlist either matches the ground truth or it does not.
3/ The leaderboard: GPT-5.5 solved the most tasks outright, 6 of 25, but earned no partial credit. Gemini 3.5 Flash leads on mean reward because it solved 5 fully and got 3 partials.
5/ It also showed real analog-design reasoning: deriving a resistor value from a feedback equation, snapping it to the E96 series, then checking a datasheet to confirm.
6/ The same capability cliff remains. Design-from-scratch tasks still time out inside the 200-step budget. The model has the domain knowledge, but long GUI execution is still where runs drift.
9/ Read the full write-up:
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-computer-use-gemini-3-5-flash/
@swyx @GoogleDeepMind 🚄🚄🚄!
@trycua @GoogleDeepMind Gemini 3.5 Flash hitting 0.267 on CUA-Bench with native Computer Use is huge. Top score, KiCad tasks no one else solves, at Flash speed and cost. That’s a W for DeepMind 🔥
@trycua @GoogleDeepMind What surprised you when people actually used it?
@trycua @GoogleDeepMind The real litmus test is complex google sheets, not excel spreadsheets; but very complex google sheet
@trycua @GoogleDeepMind The insights are of great use. The reasoning aspect has my attention
@LLMJunky @_philschmid @trycua @GoogleDeepMind extra bullish on gemini!
@trycua @GoogleDeepMind congrats @francedot
@DynamicWebPaige @trycua @GoogleDeepMind I’m trying this today
