Atomic Chat HQ dropped a demo showing a 1-bit quantized Qwen 3.8 Flash Next model hitting 30 tokens per second on a consumer MacBook Pro M5 Max.

The media thought that was the story. It was not. The media thought the story was local hardware running consumer LLMs fast. It was

Pascual ⚡Aug 27

Voxel Pagoda Art done by Qwen 3.8 Flash

CoconutAug 27

Same prompt, five models, zero edits...

Build an animated orrery in Three.js. Three of these are Claude Opus 4.6, 4.8 and 5... running through the API. The other two are open-weight models (Qwen 3.8 and Ornith 1.5) running entirely on a laptop GPU. No cloud, no API key, no

Sean DonahoeAug 27

Cubo de Rubik 3D con Qwen-3.8-flash-next

Duration: 1 hour Tokens: 4.8M Observations: In my opencode harness, it seems that one of my agents did not write and had to start over but then got it right

JuancaAug 275