Please take a moment to visit the repo where some of your concerns are addressed on the repo card itself.
2nd; publishing all the metrics at each step would be both exhausting and worse confusing.
I don't follow what you mean by "cheap" ; as a heretic [step] you usually lose 2-4 points on some metrics.
So the comparison of "heretic" vs "non-heretic" is even STRONGER ; the fairer one would be "heretic base" to "heretic tuned" which would likely show even greater change / improvement.
Source/MLX here:
https://huggingface.co/nightmedia/Qwen3.5-9B-DS9-USS-Defiant
This is on my partner's repo.
Here is a new one just uploaded; also off the scale strong, but at 9B:
Benches are up ; beats Qwen 3.5 27B in all 7 benches AND Qwen3.6 35B-A3B.
Matches some Qwen 3.6 27B benches too.
Clocks in at over 640 ARC-C for both 8bit and 4bit.
1/3 the size almost all the firepower.
Currently waiting on finalization of "4 bit" compression for these model types to address tuning/Vram issues.
These are in progress at BNB / Unsloth.
That is really the only hold up.
Otherwise VRAM to train these sparse moes is in 70 to 100 GB range. And really slow too.
Sorry no, not at this time.
This model does not contain MTP layers ; you need to run at non-MTP.
As of this writing:
There are pipeline (issues as well as optimizations) issues still currently, and it is not widely supported in some AI Apps.
Specifically:
Ggufs:
Training is compounded by number of experts in the model, which adds a serious level of time to the training.
Even 1000 samples [small!] takes 6-12 hrs.
Consider 31B dense , same samples, 30-60 minutes.
I will add to the list; may wait for specific Heretic and/or tuned version.
I already have a 43B-A3B version running in the lab ; however tuning these sparse moe models take a lot more work/time and ahh... detail. AND a lot more VRAM!!! [can't compress these atm, so BF16 required => 100 GB+ ]
I may make a Q6 high and/or a Q8 Hybrid and/or Q8 "HI".
Imatrix does not have any affect on Q8 or BF16 ; unless the other tensors in the model are set at Q6 or lower.
A Q8 "HI" is a special case; where one or more tensors/layers are set at BF16.
Currently working with Qwen 3.5/6 35B-A3B in the lab ; learning the "quirks" ; still a ways to go.
quant arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.480,0.656,0.797,0.608,0.400,0.755,0.665
mxfp4 0.455,0.607,0.851,0.585,0.402,0.744,0.651
Quant Perplexity Peak Memory Tokens/sec
mxfp8 35.937 ± 0.525 14.80 GB 1153
mxfp4 36.746 ± 0.534 11.06 GB 1030quant arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.404,0.489,0.825,0.586,0.392,0.734,0.661
mxfp4 0.414,0.508,0.854,0.562,0.378,0.717,0.645
Quant Perplexity Peak Memory Tokens/sec
mxfp8 34.652 ± 0.502 14.80 GB 1146
mxfp4 35.203 ± 0.506 11.06 GB 1200quant arc arc/e boolq hswag obkqa piqa wino
New template
mxfp8 0.518,0.709,0.755,0.657,0.418,0.759,0.626
mxfp4 0.485,0.682,0.792,0.641,0.432,0.746,0.635
Old template
mxfp8 0.506,0.697,0.754,0.661,0.416,0.757,0.627
mxfp4 0.487,0.670,0.792,0.644,0.430,0.748,0.624mxfp8 0.461,0.599,0.779,0.630,0.406,0.766,0.629
Old template
mxfp8 0.456,0.580,0.786,0.629,0.410,0.764,0.633mxfp8 0.509,0.705,0.806,0.646,0.416,0.773,0.650
Old template
mxfp8 0.502,0.692,0.809,0.650,0.420,0.771,0.651
RE: 16-18 B ; yes, something running in the lab right now. (Gemma 4).
Also can make Qwen 3's (Version 3) moes like Llama3.2-8X3B as well ; I have some of these at my repo too.
I have built a few GPT-OSS ; and some 12B [mistral nemo] as well as mistral nemo "large" 15-17Bs...
A lot of options ;
Maybe in the future ; atm still learning/addressing quirks with these new Gemmas.
Google released three different arch structure here : "E", "MOE", and 31B dense.
Also plans to create larger Gemma 4s too ; which may work better for specific applications and/or work better period.
These are in the plans for next week.
