license: apache-2.0
base_model: DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored
pipeline_tag: text-generation
library_name: gguf
description: "Two compact no-MTP GGUFs of DavidAU Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored (TWIN-TURBO Fable Cold Fusion Heretic Uncensored, 709-L): 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
• twin-turbo
• fable
• cold-fusion
  language:
• en
• multilingual

Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-BUDGET-GGUF

Two compact GGUF files of DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored (TWIN-TURBO Fable Cold Fusion Heretic Uncensored, 709-L arc class tune 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-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-GGUF family. They share the same byte-identical 448-tensor native NVFP4 backbone (all 64 layers attention plus MLP) 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 TWIN-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.

• BUDGET is the quality-leaning budget tier: Q3_K lm_head plus Q2_K token embedding. Single-16GB friendly.
• STARVED is 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'm working on over on Ko-fi. Follow along there to keep up with new releases and the work in progress. If you'd 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

Both contain 1,107 tensors: the 448-tensor native NVFP4 backbone (byte-identical across the whole TWIN-TURBO family, verified per-tensor SHA-256) plus 657 F32 norms, scales and 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 recent llama.cpp build with NVFP4 support 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. That is the smallest possible way to serve this TWIN-TURBO tune on a single 16 GB Blackwell card.

How this was made

At a high level, the steps were:

1. Self-quantized the BF16 base model to compressed tensors NVFP4 (W4A16, all Linear targets, vision, linear attention, lm_head and MTP kept in BF16, no calibration).
2. Converted the NVFP4 checkpoint to GGUF.
3. Built each file over the shared 448-tensor NVFP4 backbone, varying only lm_head and token embedding precision per file.
4. Stripped the 15 MTP tensors, leaving 1,107 tensors per file with block_count 64 and nextn_predict_layers 0.
5. Verified per file NVFP4 tensor count and backbone byte identity against the MTP family, and patched GGUF KV for name, description and license without changing tensor data.

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

Vision works. The TWIN-TURBO 709-L 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-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored or 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, text-only). That frees the projector's about 1 GB and is another lever to fit bigger context.

Loading

• Requires a recent build with NVFP4 (GGML type 40) CUDA kernels and sm_120 support (Blackwell).
• No MTP speculative path. These files have no MTP head.
• Native context is 262,144 tokens.
• For vision input, pair with 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 direct GPU runs used a diverse synthetic English payload (about 75 kB). Endpoint numbers below are end to end and include endpoint overhead, do not compare them to direct decode.

Smoke: PASS (2 of 2)

Prompt What is a black hole? via single turn sampling. Both tiers answered coherently with no repetition collapse (BUDGET distinct-5 0.9936, STARVED 0.9932).

Perplexity (PPL)

Direct perplexity on the 75 kB diverse payload, single GPU, same chunks as the 6-tier MTP family:

BUDGET 3.25 lands at the top edge of the MTP family range (3.19 to 3.25 on the same payload). STARVED 3.40 sits about 0.15 above BUDGET, the expected Q2_K floor cost.

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

Single 16 GB class card with full GPU offload, three runs per tier:

Prefill about 2697 tok/s and decode about 47 to 48 tok/s on a single 16 GB card, flat across both tiers and in line with the MTP compact tiers. Both fit a single 5070 Ti class card with full GPU offload.

Endpoint check (naive, one-at-a-time)

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

Both tiers ran repetition clean: content distinct-5-gram 1.0 with no adjacent dupes. BUDGET reasoning showed mild source citation repetition (0.94, max repeat 3) which per runbook is not a loop signature.

What to take away, naive framing

• Single run only, one large synthetic prompt, endpoint overhead included, no warmup average. Do not compare endpoint tok/s to direct numbers and do not compare across hardware.
• Both files fit a single 16 GB card (13.6 to 13.7 GiB) and decode around 47 to 48 tok/s direct. Gateway round trips finished stop at about 105 to 108 tok/s end to end.

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 and Qwen team for the base model, Qwen/Qwen3.8-27B (Apache 2.0): 27B dense, 64 blocks, Gated DeltaNet plus Gated Attention hybrid, native vision language, 262,144 token context, MTP head.
2. DavidAU for the tune itself, Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored (Apache 2.0): the TWIN-TURBO Fable plus Cold Fusion plus Heretic and Uncensored stack that this family converts.
3. Unsloth, whose trainers and systems power the underlying training methods.
4. This repo's author for the GGUF conversion and the tier ladder only, including stripping the MTP head for these two files.

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-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-BUDGET.gguf (14.72 GB)
• Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-STARVED.gguf (14.59 GB)

SHA-256

6655b4b23d16939aa44f7655509228d3c4689b674bd0e081f7a9b2054b5a4fa3  Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-BUDGET.gguf
949dfffeadd45f52e541bac07cba13c61d9ceb062519dccf091bc757599f3ef8  Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-STARVED.gguf

See also

The full MTP family (VERY-LOW, LOW, MEDIUM, MID-HIGH, HIGH, VERY-HIGH, all with the speculative head baked in) lives in esatapedico/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-GGUF. The budget tiers share its byte-identical 448-tensor NVFP4 backbone with the MTP block removed.

License

apache-2.0 (inherits from Qwen base and DavidAU tune). general.license = apache-2.0 is set inside every GGUF. general.name matches the tier name, qwen35.block_count 64, qwen35.context_length 262144.

Card 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.