license: apache-2.0
language:

• en
  base_model:
• DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored
  base_model_relation: quantized
  library_name: gguf
  pipeline_tag: image-text-to-text
  tags:
• gguf
• nvfp4
• qwen3_8
• qwen3.8
• blackwell
• mtp
• speculative-decoding
• vision
• multimodal
• llama.cpp
• lm-studio
• heretic
• uncensored
• finetune
• Cold Fusion
• tool-use
• conversational

⚡ Qwen3.8-27B TWIN-TURBO Fable Cold Fusion 709 ULTRA HERETIC Uncensored, NVFP4 GGUF

Follow me on X @BennyDaBall_OG !

This is a native NVFP4 GGUF of
DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored,
made for LM Studio and any other llama.cpp runtime on Blackwell GPUs.

The model is DavidAU's work. The training, the Cold Fusion / GAIN tuning,
the TWIN-TURBO short-reasoning behavior, the five thinking and five instruct
modes, the Heretic de-censoring, and the chat templates are all his. Nothing was
trained or edited here. This repo only changes the storage format: a fresh
one-pass quantization from his BF16 safetensors, with the MTP head and the
vision tower kept.

For everything about how the model behaves and how to drive it, read DavidAU's
docs:
NEO MTP GGUF repo and usage guide.
If you want K-quants or imatrix quants for non-Blackwell hardware, get them
there.

🧠 What is this?

• GGUF, ready for LM Studio and llama.cpp. One file for text, one optional
  file for vision. No Python, no vLLM.
• Native NVFP4 backbone. The large attention, Gated DeltaNet, and MLP
  matrices in all 64 text layers are GGML tensor type 40 (nvfp4), NVIDIA's
  4-bit floating point format with hardware kernels on Blackwell.
• MTP head kept in BF16. The GGUF has 65 blocks. blk.64 is the model's own
  multi-token-prediction layer, so the runtime can self-speculate with no
  external draft model. Output head and token embeddings are BF16 too.
• Exact-source vision. The BF16 mmproj was converted from the same DavidAU
  checkpoint, not borrowed from another Qwen3.8 model.
• DavidAU's chat template embedded unmodified (sha256 4e52a372..., verified
  byte-identical after conversion). His {REASON:...} mode switches work as
  shipped. His alternate v2 template is included as a separate file.
• 262,144 native context. 18.34 GiB instead of 51.8 GiB of BF16.

📦 The files

f4b663ec672a25e9ca5b62450822ca9b6c8cf0b56f49ce33fc050bd6813f9424  *Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored-NVFP4.gguf
dfe15f4cdd7cd6d073182e81b666d7f335890f79aff589666a106b64db32cfec  *mmproj-Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored-BF16.gguf
4e52a37225304e0fdfbd1e48594c4808f554d6e74ed2e35b954bf85dd72623f5  *chat_template.jinja
ac0b3a1299ea84e92663f0e5e85bc0d6e6d0e94f9a47dc6e48d62d97424c235a  *chat_template-tturbo-v2.jinja

🛠️ Requirements

• An NVIDIA Blackwell GPU for the native NVFP4 kernels. Built and measured on
  one 32 GB RTX 5090 (sm_120).
• A current llama.cpp or LM Studio runtime. NVFP4 tensors and the embedded
  MTP path are both recent additions. Tested working:
  • official llama.cpp release b11026 (Windows CUDA 13.4 build)
  • LM Studio with the llama.cpp CUDA 12 runtime 2.41.0
• A llama.cpp build from 2026-09-01 loaded the model fine but crashed when
  --spec-type draft-mtp was enabled. If MTP asserts at load, update.

🚀 Running it

LM Studio

1. Download the main .gguf into your LM Studio models folder. Put the
   mmproj file in the same folder if you want image input. LM Studio pairs
   them automatically.
2. Load it like any other GGUF. The chat template is read from the file.
3. For the speed boost, enable Draft MTP speculative decoding in the load
   settings with max draft tokens 3 and min continue probability 0.2.
   From the CLI:

lms load qwen3.8-27b-twin-turbo-fable-cold-fusion-709-ultra-heretic-uncensored-nvfp4 \
  --gpu max --context-length 32768 \
  --speculative-draft-mtp \
  --speculative-draft-max-tokens 3 \
  --speculative-draft-min-continue-probability 0.2

The model thinks by default. To switch modes from inside the chat, use
DavidAU's tags at the start of a message, for example {REASON:imedium} for a
direct answer with no thinking block or {REASON:low} for brief thinking. The
tag stays in effect until you change it. This was tested in LM Studio and works
through its template engine.

llama.cpp, text with embedded MTP

llama-server \
  --model Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored-NVFP4.gguf \
  --no-mmproj \
  --ctx-size 131072 \
  --flash-attn on \
  -ctk q8_0 -ctv q8_0 \
  --spec-type draft-mtp \
  --spec-draft-n-max 3 \
  --spec-draft-p-split 0.2 \
  --jinja \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.0

llama.cpp, vision

Replace --no-mmproj with
--mmproj mmproj-Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored-BF16.gguf.

Agent harnesses and tool calling

Tool calling works on the embedded default template. Like the stock Qwen3.8
template, it rejects a system message that arrives after the first turn, and
some agent harnesses send those. For that case use DavidAU's v2 template, which
accepts them and keeps the same {REASON:...} modes:

  --chat-template-file chat_template-tturbo-v2.jinja

Both templates were tested here with a tool call and a tool-result follow-up.

🎛️ Modes and samplers (from DavidAU's docs)

Switch with a tag in your message or with reasoning_effort / enable_thinking
over the API:

DavidAU's suggested settings: thinking mode temp 1.0, top_p 0.95, top_k 20,
instruct mode temp 0.7, top_p 0.8, top_k 20, presence penalty 1.5. For MTP he
recommends temperature at or below 1.0 and repetition penalty 1.0 (off). See his
guide for the full details.

📈 Measured on one RTX 5090

Official llama.cpp b11026, 131,072 allocated context, Q8_0 K/V cache, flash
attention, one slot, greedy, two runs per workload. VRAM is total board usage
and includes about 3.0 GiB of desktop apps.

Embedded MTP was +87.5% mean decode speed, and the generated text was
identical with it on and off.

LM Studio (runtime 2.41.0, 32,768 context, same greedy code prompt, identical
output text both ways): 70.8 tok/s with MTP off, 137.4 tok/s with Draft
MTP at max 3 / probability 0.2. Load time was about 10 seconds.

These are short generations on one machine, not a benchmark suite. Acceptance
depends on the workload and the sampler: one sampled creative-writing prompt
with presence penalty 1.5 accepted only 31% of drafts, and MTP gave no gain
there. Code and reasoning is where it pays.

Every runtime was also checked by reading the output, not only the counters:
generated Python that ran and passed its own asserts, correct arithmetic and
word problems, clean prose, a correct read of three values from a test image,
and no looping or junk tokens.

🧬 The conversion, briefly

Fresh one-pass quantization from DavidAU's BF16 checkpoint at revision
2d28f63aa79cab53e19fb26139d69eb22940d9c2. Not a requant of another GGUF.

1. Downloaded the source and verified all 27 files against the Hub checksums.
   The source has 1,199 BF16 tensors, including all 15 MTP tensors and the 333
   vision tensors. mtp.fc.weight lives in DavidAU's separate
   model-mtp-restored.safetensors and is picked up through the index.
2. Converted to a BF16 parent GGUF with an MTP-aware llama.cpp converter
   (commit 5ecbe1ac), which maps mtp.* to blk.64.
3. Quantized once with llama-quantize, an importance matrix, and a complete
   per-tensor type map. The q8_0 argument is only a fallback, the map decides
   every tensor.
4. Converted the vision tower to a BF16 mmproj from the same checkpoint.
5. Rewrote the container metadata only, to remove a local file path. All 866
   tensors were verified byte-identical before and after.

python convert_hf_to_gguf.py "$SOURCE" --outfile parent-bf16.gguf --outtype bf16

llama-quantize \
  --imatrix Qwen3.8-27B-NVFP4-imatrix.dat \
  --tensor-type-file Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored-NVFP4-tensor-types.txt \
  parent-bf16.gguf final-nvfp4.gguf q8_0 32

python convert_hf_to_gguf.py "$SOURCE" --outfile mmproj-bf16.gguf --outtype bf16 --mmproj

Result: 866 tensors, 65 blocks, one NextN layer. 448 NVFP4 (the large
backbone matrices), 408 F32 (small state, conv, norm, and scale tensors),
10 BF16 (output head, token embeddings, and the eight large MTP matrices).
18,771.52 MiB at 5.76 bits per weight. The projector is 334 tensors, 110 BF16
and 224 F32.

The importance matrix is architecture-matched Qwen3.8-27B calibration (200
chunks, 496 entries). It was not computed on this fine-tune. It only guides
rounding, and both it and the tensor map are included here so the recipe is
fully inspectable.

⚠️ Notes

• This is an uncensored model. DavidAU reports 6/100 refusals against 86/100
  for untuned Qwen3.8-27B, at KL divergence 0.0397. Those are his numbers and
  were not re-measured here. You are responsible for how you use it.
• NVFP4 targets Blackwell. For other GPUs or CPU use DavidAU's regular GGUFs.
• Quantization always costs something. No quality benchmark was run on this
  file beyond the checks described above.

🙏 Credits

• DavidAU for the model, its training
  project, its chat templates, and the usage guide. If you like this model,
  like and follow his repos.
• Alibaba / Qwen for Qwen3.8-27B.
• The llama.cpp maintainers, and the Heretic project used upstream.

Apache-2.0, matching the upstream model. "Qwen" is a trademark of Alibaba and is
used only to identify the upstream model. This repository is not affiliated with
or endorsed by Alibaba or DavidAU.

Quantized locally by BennyDaBall.

Follow me on X @BennyDaBall_OG !