What are people getting w/ Qwen-3.8-27B on M5 Pro with 24gb?
I think we might be close to 2x'ing (this: with the iPhone handling spec decode and mac handling verification. probably some room to improve here. Just glad that my 24gb machine might not be
GLM-5.3 Flash and Qwen 3.8 Flash same instructions in Topview Canvas.
1 same image. Same Seedance 2.5. The only difference is the LLM that wrote the prompt. ・The only thing I asked LLM to write was "one prompt." The generation conditions are fixed here and the same conditions ・480p / 16:9 / 15 seconds / One shot / No retry
GPT-6 Astra just built a 3D evolving ecosystem in a single HTML file.
One shot. No errors. Raw WebGL and GLSL. No libraries. Under 30 minutes. I wanted to test something specific. Not coding ability. Emergence. Evolution is one of the most complex systems you can ask a model to
Interesting, I’m using my own harness too, renaming the tools to make their purpose more explicit helped a lot.
Just ran your exact same prompt, on Qwen 3.8 27B Q4, with reasoning set to xhigh. 20m10s 1.93M input tokens 66.9K output tokens 24 model calls Here’s the result 👇
You have reached the end of the archive
All of qwen38