Play along with this 0-shot from Alibaba_Qwen Qwen 3.8 27B at FP16 on vllm_project strapped up to NousResearch Hermes Agent👇

Digital SpaceportAug 146

I spent time testing Qwen 3.8 27B.

First impressions: - 47 tokens/sec with MTP. - It is a real improvement over 3.6 in quality on the voxel test. All the elements are properly grounded, nothing upside down, and and nice movement in the water. - It thinks ALOT. Nearly unusable on

Jackson AtkinsAug 147

And here are Qwen 3.8 27b Q8 websites definitely better than the NVFP4.

Also getting really good game generation with Q8. Q8 is a win.

Loktar 🇺🇸Aug 1436

Here's Qwen 3.8 27B 4-bit running on Asus GX10 (=DGX Spark).

Sure, it's not GPU speeds but quite okay with multi-token prediction enabled. Plus, many of my workflows are 'plan first for an hour, then push a button, return in x hours', and for that the speed is more than enough.

Topi SantakiviAug 141