There's been a lot of hype about Ternary Bonsai2 but fw demonstrations on how it actually compares to the full 16 bit version of Qwen 3.8 27B.
Since I have my own quantized version of that model thatI'm testing, ~12gb file size, I decided to throw Bonsai2 in there also to get a
Now I tested the difference between Qwen 3.8 27B local vs GLM 5.3 both on Droid
There is a difference in quality and being able to follow the images and attention to detail more closely. Clearly GLM 5.3 is ahead in quality, although it took much longer to build the game.
Qwen 3.8 Omni Flash can WATCH your videos, not just read the transcript.
That means AI can finally understand what was said AND what happened on screen at the same time. Here’s how I’d test it: → Open the Qwen chat app. → Pick Qwen 3.8 Omni Flash. → Upload a video. →
I love the beautiful sadness of this
Anyway related, here is harness on a rented NVIDIA H200 141GB optimizing training of QWEN-3.8-27B on a complex reasoning task -runs a few short trials on a small sample of data, watches where the time ((per-step time, nvtop, py-spy) -applies
You have reached the end of the archive
All of qwen38