My AI agent just finished a full Duolingo lesson on a real iPhone.
And it didn't burn the most expensive model on every step. That's Handsy (ESP32 Agent): a small ESP32 board that gives an agent a real iPhone. 700+ people have installed it via TestFlight, and the repo just
I gave all three of my local models the same prompt to build a neon rain alley in three.js.
GLM 5.3 Flash ran at 95 tok/s and DeepSeek V4.1 Flash at 127 on my 4 box Spark ring, and Qwen 3.8 27B did 42 on my Mac Studio while it was busy with other work. Which one built it best?
1/ I have been accused several times that local LLMs are rubbish and that ONLY Opus 5.5 works, especially in web development.
Everything else is completely useless. Ok, so let's simulate a V8 in 3D with camera panning, engine cutting, exploded view, etc... Yes, they are
Good guy! The perfect partner for Fable 5.2 is actually the Qwen 3.8 27B running locally!
Atomic Agent made a comparison with a three-dimensional fairy tale book: Fable only divided tasks and assigned tasks, and left all the specific work to the local Qwen. Compare the two sides: 🔸Finished product: The matching version of the hot air balloon, castle, and snow mountain is very layered, and the one made by Fable alone looks ordinary. 🔸Time consuming: one 11…
You have reached the end of the archive
All of qwen38