DavidAU

/

Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF

like 2.27k

Thank you so much ; and thank you for Unsloth.
Absolutely the best training tool on the planet.

This image tells a story I am still struggling with:

[from IRONLLM Labs]

https://blog.robai.net/27bevals/
See the full report on Fusion 711 in full detail at every level.

@DavidAU can you give us an BF16 version of this model to download? I want to convert it to ONNX. Or you just offer your model downloads in ONNX direclty, this would be also very welcomed :)

Here you go:
https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP

Here you go:
https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP

I can only see the quantized versions between 2 and 8 bit, but no ONNX or FP16 (Unquantized) version. Or is the "mmproj-F16.gguf" all i need and i need to use it altogether with the raw Qwen 3.6 27B FP16 model? Sorry iam new to this :D

The source contains raw BF16 , with all the "proj" files (images,video tensors) embedded.
Download this and you can create any quant and/or fine tune the model.

Go to the FILES tab at the repo; the other "quants/quantized" is derived from the source code.

The source contains raw BF16 , with all the "proj" files (images,video tensors) embedded.
Download this and you can create any quant and/or fine tune the model.

Go to the FILES tab at the repo; the other "quants/quantized" is derived from the source code.

Nice, now i see. Thank you very much mate :)

I can only see the quantized versions between 2 and 8 bit, but no ONNX or FP16 (Unquantized) version. Or is the "mmproj-F16.gguf" all i need and i need to use it altogether with the raw Qwen 3.6 27B FP16 model? Sorry iam new to this :D

Nothing wrong with being new to this :)

In short, the safetensors usually is F16 (floating 16) and will likely be the highest quality you can download.
GGUF is a popular format that holds all the data in a single file, and is often quanitized. Probably the most popular are 8_0, Q6K and Q4_K_M. Anything lower than Q4 the quality degrades a lot.
mmproj is a vision/audio addon, basically it will encode files you pass to the model so it can analyze it according to what it's expecting. On the back end you'll see it UU64 encoded.

The size of the model closely resembles how much memory you need. So a 27B will need 27Gb for a 8_0 model. smaller ones will take up less space. If you don't have enough VRAM then it will likely run on CPU which is very slow. So often it's better to get models that will fit well on your GPU. MOE's the A component (A3B for example) as that's the active nodes and those tend to run a lot faster even if all of it won't fit on your card.

I'm mostly been using Llama.CPP (thus oobabooga), and KoboldCPP so i almost exclusively use Q6K GGUFs. Easy to load, and overall to run with all the tools you need to start. (Well maybe not tool calling, but that's a side point as it gets improved)

The link provided you replied to was a safetensors, so it's unquanitized. If you want the quantized you need to sift through other card pages, like ones done by mradermacher, or DavidAU does his own. Going to the quanitized link will show you related based on the mode. Choose according to your hardware. If you get a larger version you can quanitize down later probably.

Go to the FILES tab at the repo; the other "quants/quantized" is derived from the source code.

now i wonder if my reply was a little overkill...

@yano2mch well i am not new to LLM use in general. My only problem was downloading models from Huggingface manually. I am using LM Studio for years, and only got my models through the internal UI. Now i needed to switch to Lemonade, since my 128GB Strix Halo machine cannot utilize the NPU + GPU together with LM Studio's llama.cpp runtimes. Lemonade natively supports FLM and ONNX (Hybdrid model runtimes), so i now needed to get the proper formatted model manually. I can fit some huge monsters in 128GB VRAM, so i am not very limited in that manner. Also i will upgrade to a dedicated homelab AI machine with 256GB as minimum goal, but i can currently only afford buying 128GB in dedicated HBM2 AI GPU's.
Thanks tho.

Update: (repeating across all 3.8 threads || Primary "Update" thread.)

Currently we are testing Qwen 3.8 27B to get core metrics and primary "human testing" to get a sense of model's
character /stats / issue(s) as well as how it compares to base Qwen 3.6 27B, Qwen 3.6 27B Fable Fusion 711 and other lab versions of "711".

As of this writing we know the pipeline is compatible with this model, as well as the new GAIN training method (AKA "COLD FUSION")
which was invented during the build of Qwen 3.6 27B Fable Fusion 711.

Fable Fusion 711 used a primitive (but effective) version of GAIN during the build process. COLD FUSION takes this to the next level
with the ability to maintain BF16 metrics across 8 bit and 4 bit at 99% (of BF16/full precision) as of this writing.

The first test model (built 14 days ago) has been released here:
https://huggingface.co/DavidAU/Qwen3.5-9B-Cold-Fusion-GAIN-v1.0-Uncensored-Heretic-NEO-MAX-Imatrix-GGUF

With Qwen 3.6 Fable Fusion "Darker Hero" (built 3 days ago) also using COLD FUSION here (ggufs/source releasing shortly):
https://huggingface.co/DavidAU/Qwen3.6-27B-V1.1-FF711-Darker-Hero-GAIN-H2.0

NOTE:

• This version has been "re-heretic'ed", and trained to reduce thinking down to 1/10 to 1/2 "qwen size".
• 3 example generations.

We published both "reg training" and "GAIN training" benches (at both 9B and 27B repos above) so you can see the differences.
They are stark.

With Qwen 3.8 27B we will shortly be testing it in the pipeline and/or using COLD FUSION on it.

Conversely GAIN will also be used on current Qwen 3.6 models, and other arch/model types.

This is a lengthy, intense and detailed process so it may take some time before we publish.

Thank you ;