---

license: apache-2.0

base_model: DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU

pipeline_tag: text-generation

library_name: gguf

description: "Two compact no-MTP GGUFs of DavidAU Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU (TURBO Fable Cold Fusion Heretic Uncensored, 735/882): BUDGET (Q3_K lm_head) and STARVED (Q2_K everything). Byte-identical 448-tensor native NVFP4 backbone, no MTP head, single-16GB friendly. 262,144 native context. Blackwell sm_120."

tags:

- gguf

- nvfp4

- qwen3.8

- qwen3.5

- blackwell

- vision

- multimodal

- llama.cpp

- turbo

- fable

- cold-fusion

language:

- en

- multilingual

---

Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-BUDGET-GGUF

Two compact GGUF files of DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU (TURBO Fable Cold Fusion Heretic Uncensored, 735/882 Heretic merge and DPO layers on Qwen3.8-27B, 27B dense hybrid with Gated DeltaNet plus Gated Attention, 262,144-token native context, native vision tower) - without the MTP speculative head.

These are the budget members of the esatapedico/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-GGUF family: they share the same byte-identical 448-tensor native NVFP4 backbone (all 64 layers attention plus MLP, 13.69 GB of NVFP4) as the MTP tiers, but the MTP draft block has been stripped and the extra tensors pinned to the smallest types. No drafter, no speculative decoding, just the model at the lowest possible footprint. Pairs with the original Qwen3.8 vision projector.

The goal is to let 16 GB VRAM users benefit from native NVFP4. These are the smallest possible ways to serve this TURBO tune on a single 16 GB Blackwell card. There are, of course, accuracy trade-offs vs the larger MTP tiers, plus the lack of MTP. Like every compact tier of the family, the NVFP4 backbone is preserved byte identical, and here the head tensors are pinned to the smallest types. You can also fit TURBO 735-882 in 16 GB through non-NVFP4 variants if you prefer a different quantization.

• BUDGET - the quality-leaning budget tier: Q3_K lm_head plus Q2_K token embedding. Single-16GB friendly.
• STARVED - the absolute floor: Q2_K lm_head and Q2_K token embedding. Smallest possible variant of the family.

Follow along & support

I post updates on new conversions, benchmarks, and what I am working on over on Ko-fi. Follow along there to keep up with new releases and the work in progress. If you would like to support more of it, a coffee is always welcome. I do this on consumer hardware and like seeing how far it goes. More is on the way.

☕ ko-fi.com/esatapedico. Updates, work-in-progress, and an optional coffee.

The two files

| File | Size (decimal GB) | lm_head (output.weight) | token_embd | MTP head |

|---|---|---|---|---|

| Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-BUDGET.gguf | 14.72 GB | Q3_K | Q2_K | - (none) |

| Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-STARVED.gguf | 14.59 GB | Q2_K | Q2_K | - (none) |

Both contain 1,107 tensors: the 448-tensor native NVFP4 backbone (byte-identical across the whole TURBO family, verified per-tensor SHA-256) plus 657 F32 norms/scales/gates plus the two head tensors. qwen35.nextn_predict_layers=0 and qwen35.block_count=64 (the MTP block is gone), so they load in any llama.cpp build without the draft-mtp path.

Why these exist

The MTP family is built around a byte-identical NVFP4 backbone; the MTP draft head costs VRAM and compute on every pass. If you do not use speculative decoding, the stripped versions free that head entirely and shave the extra tensors down to the cheapest types - the smallest possible way to serve this TURBO tune on a single 16 GB Blackwell card.

Vision works. The TURBO 735-882 tune leaves the original Qwen3.8 vision tower untouched. Pair either GGUF with the mmproj-BF16.gguf from the MTP repo or the base model DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU / unsloth/Qwen3.8-27B-GGUF projector via --mmproj. If you need the absolute minimum VRAM footprint on a single 16 GB card, you can also load the model without the projector (no --mmproj, i.e. text-only) - that frees the projector's about 1 GB and is another lever to fit bigger context.

Usage

llama.cpp / llama-server

llama-server \
  --model Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-BUDGET.gguf \
  --mmproj mmproj-BF16.gguf \
  --ctx-size 262144 \
  --flash-attn on \
  --temp 0.7 --top-p 0.95 --top-k 20

• Requires a recent llama.cpp with NVFP4 (GGML type 40) CUDA kernels and sm_120 support (Blackwell).
• No --spec-type draft-mtp - these files have no MTP head.
• For vision input, pass --mmproj mmproj-BF16.gguf (from the MTP repo or base).
• Qwen3.8 official sampling presets: thinking mode temp 1.0 / top_p 0.95 / top_k 20; instruct mode temp 0.7 / top_p 0.80 / top_k 20. Pick per use case.

Benchmarks (naive, single run, not comparable across setups)

These are rough sanity checks to confirm the files load and generate, not a formal benchmark. Method and hardware are noted so you can interpret them in context. Results will vary with hardware, sampling and context length.

All runs used direct GPU via llama.cpp tools, with a diverse synthetic English payload (about 75 kB, 28 chunks times 512 context). LocalAI gateway numbers are end to end and include gateway overhead, don't compare them to llama-bench decode.

Smoke: 2 of 2 coherent

Prompt What is a black hole? via single turn sampling, each tier produces a natural coherent completion with no repetition or truncation. Both BUDGET and STARVED passed.

Perplexity (PPL)

Direct llama-perplexity on the 75 kB diverse payload, single GPU, same chunks as the 7-tier MTP family:

| Tier | PPL (Final estimate) |

|---|---|

| BUDGET | 3.3410 +/- 0.07288 |

| STARVED | 3.4560 +/- 0.07618 |

BUDGET is identical to VERY-LOW in the MTP family (both Q3_K lm_head plus Q2_K token embedding). STARVED is the highest in the whole family at 3.4560, about 0.11 higher than VERY-LOW, as expected from Q2_K on both heads. For reference the MTP family spans HIGHEST 3.2367 (best, fidelity-max 256 NVFP4 plus Q8_0) to VERY-LOW 3.3410, with LOW 3.2761, MEDIUM 3.2858, MID-HIGH 3.2903, HIGH 3.2851, VERY-HIGH 3.2835. STARVED sits about 0.22 above HIGHEST.

Speed (llama-bench, pp512 prompt processing, tg128 generation, tok/s)

Single 5070 Ti (16 GB class) with full CUDA offload, three runs per tier:

| Tier | VRAM (llama-bench) | pp512 tok/s | tg128 tok/s | Notes |

|---|---|---|---|---|

| BUDGET | 13.70 GiB | 2753.10 +/- 249.27 | 47.51 +/- 0.06 | fits, full GPU |

| STARVED | 13.58 GiB | 2762.68 +/- 235.81 | 48.05 +/- 0.13 | fits, full GPU |

Both fit comfortably on a single 5070 Ti 16 GB card at full GPU, about 13.6 GiB reported plus KV. Prefill around 2750 tok/s and decode 47 to 48 tok/s, flat versus the MTP compact tiers VERY-LOW 2748/47.30, LOW 2728/46.77, MEDIUM 2704/45.55, MID-HIGH 2714/45.55, which share the same 448 backbone. Larger MTP tiers HIGH 16.59 GiB at 34.89/6.32, VERY-HIGH 17.99 at 34.44/6.12, and HIGHEST 19.81 at 35.40/4.73 exceed single 16 GB and fall back to CPU on this bench, they need dual-GPU split or 24 GB.

LocalAI gateway check (naive, one-at-a-time)

Each tier was sent one large request with max_tokens=20000 and finish_reason=stop for both. These are end to end gateway timings, not pure decode, so treat as sanity and responsiveness, not a formal benchmark.

| Tier | request_seconds | tokens per sec (20000 / s) | finish_reason | reasoning chars | content chars |

|---|---|---|---|---|---|

| BUDGET | 178.84 | 111.83 | stop | 1162 | 793 |

| STARVED | 303.68 | 65.86 | stop | 1800 | 10811 |

Both finished with finish_reason stop and distinct 5-gram ratios near 1.0 with no repetition collapse. Tokens per sec is 20000 divided by request_seconds and reflects the whole gateway round trip for this large prompt, not llama-bench decode.

For reference, the MTP family on the same gateway method measured VERY-LOW 661.94s at 30.21 tok/s, LOW 370.89 at 53.92, MEDIUM 374.34 at 53.43, HIGH 507.82 at 39.38, VERY-HIGH 505.75 at 39.55, MID-HIGH 422.28 at 47.36, and HIGHEST 416.45 at 48.02. All gateway runs 20000 tokens, finish stop, one-at-a-time, naive single run.

What to take away, naive framing

• Single run only, one large synthetic prompt, gateway overhead included, no warmup average. Don't compare gateway tok/s to direct llama-bench numbers and don't compare across hardware.
• BUDGET and STARVED both fit single 16 GB with decode around 47 to 48 tok/s on direct bench. Gateway round trips for this prompt were faster than the MTP compact tier average, but that reflects prompt length and gateway caching, not a decode win.

Attribution & provenance

This is a derivative work built entirely from existing Apache-2.0 artifacts. Nothing here was trained or fine-tuned. Credit belongs to:

1. Alibaba / Qwen team for the base model, Qwen/Qwen3.8-27B (Apache-2.0).
2. DavidAU for the tune itself, DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU (Apache-2.0): the TURBO 735-882 Fable plus Cold Fusion plus Heretic/Uncensored DPO stack.
3. Unsloth for the NVFP4 quantization approach and the vision projector unsloth/Qwen3.8-27B-GGUF (Apache-2.0, mmproj-BF16.gguf).
4. This repo's author for the GGUF conversion, the tier splicing, and stripping the MTP head.

The NVFP4 tensors are native GGML type 40, preserved from the source checkpoint through GGUF conversion with no re-quantization round trip.

Repository contents

• Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-BUDGET.gguf (14.72 GB)
• Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-STARVED.gguf (14.59 GB)

SHA-256

b938c42d0d5f5c85641360c28bfece301eb8179f2b29ff1d31728095e9e19132  Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-BUDGET.gguf
6a050fc09f565c54eab8336d070b4e5e4ff0a3f92fa2cd1d5ae290b389a9c10d  Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-STARVED.gguf

See also

The full MTP family (VERY-LOW / LOW / MEDIUM / MID-HIGH / HIGH / VERY-HIGH / HIGHEST, all with the speculative head baked in) lives in esatapedico/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-NVFP4-GGUF. The budget tiers share its byte-identical 448-tensor NVFP4 backbone with the MTP block removed.

Benchmark summary for that family: VERY-LOW 3.3410 ppl 2748/47.30, LOW 3.2761 2728/46.77, MEDIUM 3.2858 2704/45.55, MID-HIGH 3.2903 2714/45.55, HIGH 3.2851 34.89/6.32 CPU fallback, VERY-HIGH 3.2835 34.44/6.12 fallback, HIGHEST 3.2367 best at 35.40/4.73 fallback single but 48.02 tok/s via dual-GPU gateway at 416.45s. Gateway for that family was VERY-LOW 661.94s 30.21, LOW 370.89 53.92, MEDIUM 374.34 53.43, HIGH 507.82 39.38, VERY-HIGH 505.75 39.55, MID-HIGH 422.28 47.36, HIGHEST 416.45 48.02. All gateway runs 20000 tokens, finish stop, both GPUs default, one-at-a-time, naive single run.

License

Apache-2.0, identical to every upstream artifact. The base model license governs; GGUF conversion and quantization are transformations, not new training. When redistributing, please retain attribution to Qwen (Alibaba), DavidAU, and Unsloth as above.

"Qwen" is a trademark of Alibaba. Trademarks are used here only to identify upstream models; this repository is not affiliated with, sponsored by, or endorsed by Alibaba, DavidAU, or Unsloth.

Note on this card

This model card was written by AI assistance at the request of the repository author, who did the engineering. As with any AI-generated text, there may be errors; please verify anything important (hashes, sizes, commands) against the file itself before relying on it.