Those are the recommendations, but I used an actual measured sensitivity. I sanity check each model with real agentic coding loops before uploading. Are you seeing problems with that model?
l tried a few tests with the "quality" version and found it performing worse than original 3.6 35B A3B I'm using right now. Too many bugs, syntax errors, wrong initialization order, etc. Just compared againts recently released Qwen 3.8 by unsloth and found there
ssm.alpha IQ4_XS ssm.beta IQ1_M
GPT searched for source code and found:
beta = ssm_beta · x beta = sigmoid(beta)
alpha = ssm_alpha · x alpha = alpha + dt alpha = softplus(alpha) gate = alpha * A
The conclusion was, that modern Unsloth's decision to keep alpha in higher quants is reasonable.
Kat is one I mostly produced to demonstrate that the Qwen MTP worked with a reasonable acceptance rate. (Kat coder is published without MTP). I'll take a look at it.
So I did rebuild with F32 and my benchmarks improved (The old version got a perfect score on my programming benchmarks, but the new version took half as many tokens for the same result.) Good catch.
I'm also rerunning an MTP settings sweep but expect n=2 will be optimum like it is for other Qwen 3.6 derivatives with the recommended temperature. If you run it cooler the MTP helps more.
Update - Nope n=3 was about 1% better on my sweep.
These run great on my 16GB rtx 4060 with offloading and of course they fit fine on the dgx spark. I have even run them CPU only and they're slow but usable. These are typically bandwidth bound, so smaller => faster; bigger => slightly smarter.
@Myric, thanks for rebuilding and alpha/beta role confirmation. I usually try hard test prompt:
Write a single-file HTML/JS/CSS Windows-style desktop application with six apps, a start menu, a task manager, and a console. No external dependencies. The whole desktop and each application must be ui friendly ,beautiful, fully functional, meaningful and useful, including all tabls and settings. Output to desktop.html
It's hard for model to mix html, css, javascript in a single 1500 lines file.
