Had Opus AMDify the nvidia Qwen 3.8 27b tips I’d seen floating around.
Got the 7900xtx up from 32t/s to 56t/s. Q4_K_XL safely with 80k context. Write up below.👇
Someone made a public Qwen 3.8 27b for free!!
Took me 30 seconds to get it into my codex. If you haven't paid for any subs or ran out of usage, this plus a couple of the free models on opencode is a real setup. Start slow and grind your way up. shoutout victor for putting it
Qwen-3.8-27B NVFP4 (weight 21.3G), MTP is not enabled, the throughput speed on DGX Spark is >10tok/s
Feel the speed👇🏻 I will turn on MTP later for comparison.
Gemini 3.7 Flash test against Qwen 3.8 and Grok 4.6 and GLM 5.3;
Same recipe, 4 models: a harvester and farmers working. to be honest Qwen 3.8 did exactly what was asked of it. Grok 4.6 failed to create textures. Which one do you think worked better?
You have reached the end of the archive
All of qwen38