ABSURD. Qwen 3.8 27B is cruising locally on an RTX 4060 with just 8GB of VRAM.
That’s a 64k context window courtesy of Unsloth’s brand-new IQ4_XS quant, and the whole thing weighs in at 14.6GB on disk. Prefill lands around 150 tok/s; decode hovers at ~5 tok/s with native MTP.
Japanese Tea Garden Bench - Qwen 3.8 Flash Next
Just lucky, I guess. This one took 6 hours and several million tokens, at medium effort. I would say good job, but the z-fighting is prominent, and it might never have finished had I not stopped it.
This is my personal experience with Qwen 3.8 27B in Hermes versus Grok 4.6 in Grok Build.
I made the subtitles semi-transparent so it should be easier to see.
I made the subtitles semi-transparent so it should be easier to see. Completion Beatles style composition with Vox Score → Cover with Suno → Import into LaViale → Analyze with Whisper → Direct with Qwen 3.8 → Krea 2 +… I made the subtitles semi-transparent so it should be easier to see.
You have reached the end of the archive
All of qwen38