Nice
enjoyed the read nice one also seeing this being put out there Qwen3.8-120B/51B/A6B MoE is coming out in 24 hours.
What a happy day source from here https://modelscope.cn/models/Qwen/Qwen3.8-Flash-Next
Glad you enjoyed it — and thanks for the heads-up back then, because that model you flagged is out, and I ended up building the uncensored line of it:
- Flash-Next-Abliterated (bf16, refusal removed)
- Flash-Next-Abliterated-NVFP4
- Flash-Next-Abliterated-GGUF (five quants)
On the numbers you quoted — I took the arch apart (write-up in ARCHITECTURE.md on the abliterated repo, first written account of qwen4_exp): it's ~180 B total once you count the ~51 B PLE n-gram table you saw, with ~7 B active per token (the "A6B" headline), 512 fused experts, and — relevant to this survey — it ships a real MTP draft head. So it's squarely in scope: the NVFP4 above has that head grafted back and pinned in quantization_config.ignore, which is exactly the failure mode the survey found most repos get wrong (head present in name, dead in practice).
One honest footnote: it's a 180 B beast — great throughput from the 7 B active path, but not a 128 GB-class model in any quant that keeps quality, unlike the 27 B. Fun full circle, though: you pointed at it, and now it's abliterated and quantized.
i got over the disapointment with the 3.8 flash as i seen this the stats are insane for a 27b https://huggingface.co/DavidAU/Qwen3.8-27B-Cold-Fable-Fusion-GAIN-V1.1-732-Heretic-Uncensored-stage1 i deffenetly want to jump on tthis one when its out if yoou nvfp4 it ! top tier
Happy to — with one timing caveat I want to be straight about. I looked at that repo: the weights are up (~55.6 GB bf16), but it's still an explicit work-in-progress. The card itself says it's a WIP "until final version(s) is released," lists stages 1c/1d/1e as pending, and even carries a "not happy with this step, making adjustments" note — and the name literally ends in "stage1." It's also gated, so I'd need to be granted access before I could even pull it.
Quantizing a moving target means the NVFP4 could be stale within days, so the right move is to wait for DavidAU to mark it final, then build it — same pipeline as the GAIN one I did (MTP head grafted/restored, vision tower kept bf16). One thing I couldn't confirm while it's gated is whether all the mtp.* tensors and the vision weights are complete; there's a model-mtp-restored.safetensors in there, which is a good sign, but I'll verify at tensor level once it's final and open.
So yes — it's on my list. I'll watch for the finished release and post in this thread when the NVFP4 is up. Ping me if you spot DavidAU flipping it to final before I do.