license: apache-2.0
base_model: DavidAU/Qwen3.8-27B-Ultimate-Heretic-Uncensored
pipeline_tag: text-generation
library_name: gguf
description: "Two compact no-MTP GGUFs of DavidAU's Ultimate-Heretic Turbo-Brilliance-50x-AGI-NEO tune of Qwen3.8-27B (NVFP4): BUDGET (Q3_K lm_head) and STARVED (Q2_K everything). Byte-identical 448-tensor NVFP4 backbone, no MTP head, single-16GB friendly. Blackwell sm_120."
tags:
- gguf
- nvfp4
- qwen3.8
- blackwell
- ultimate-heretic
- llama.cpp
language: - en
- multilingual
Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-BUDGET-GGUF
Two compact GGUF files of DavidAU/Qwen3.8-27B-Ultimate-Heretic-Uncensored, DavidAU's Ultimate-Heretic Turbo-Brilliance-50x-AGI-NEO tune of Qwen3.8-27B — without the MTP speculative head.
HF repo: esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-BUDGET-GGUF
These are the budget members of the esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-GGUF family: they share the same byte-identical 448-tensor NVFP4 backbone as the seven 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.
The goal here is to let 16 GB VRAM users benefit from native NVFP4. These are the smallest possible ways to serve this tune on a single 16 GB Blackwell card. There are, of course, accuracy trade-offs vs the larger tiers, plus the lack of MTP.
BUDGET— the quality-leaning budget tier:Q3_Klm_head +Q2_Ktoken embedding. Single-16GB friendly.STARVED— the absolute floor:Q2_Klm_head andQ2_Ktoken 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 my site. Follow along there to keep up with new releases and the work in progress. If you'd like to support more of it, you will find the options there. I do this on consumer hardware and like seeing how far it goes. More is on the way.
🌐 esatapedico.com. Updates, work-in-progress, and ways to support.
The two files
| File | Size | lm_head (output.weight) |
token_embd | MTP head | SHA-256 |
|---|---|---|---|---|---|
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-BUDGET.gguf |
14,723,056,960 bytes (14.72 GB) | Q3_K |
Q2_K |
— (none) | 6b04d202f1583a4384ae3c610699e624827685a18ff280fac9e4779b063adb12 |
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-STARVED.gguf |
14,593,930,496 bytes (14.59 GB) | Q2_K |
Q2_K |
— (none) | 7b79c749b95b9b91f6a77c6c313c92325bfc1d35d29348e3bb053faec20310dc |
Both contain 1,107 tensors: the 448-tensor NVFP4 backbone (byte-identical with the seven MTP tiers, verified per-tensor SHA-256) + 657 F32 norms/scales + 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 don't 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 tune on a single 16 GB Blackwell card.
Vision works. Pair either GGUF with the mmproj-BF16.gguf (b02e6393f72a69e2358388fbeffddc52b04898933a15d44de9ff3eb08f6099e0, 927,608,768 bytes) from the MTP repo 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 ~1 GB and is another lever to fit bigger context.
Measurements
Single-turn smoke: both files generate coherent text with a natural stop and no repetition-loop signature (BUDGET 27.9 t/s decode, STARVED 30.4 t/s).
Perplexity on a source-diverse corpus (lower is better):
| Tier | PPL |
|---|---|
| BUDGET | 3.3414±0.07566 |
| STARVED | 3.5139±0.08154 |
BUDGET matches its VERY-LOW parent to the digit, as expected — stripping the unused MTP block changes nothing the perplexity pass reads.
Task quality (unified suite gate)
MMLU letter-logprob subset e1aee941… (n=1000, thinking off) and GSM8K (n=200, thinking on), same flags as the seven-tier ladder:
| Tier | MMLU (n=1000) | GSM8K (n=200) |
|---|---|---|
| BUDGET | 83.29% | 0.93 |
| STARVED | 82.42% | 0.935 |
BUDGET reproduces its VERY-LOW parent exactly. Every suite verdict FAILs only on the absent KLD anchor, as with the MTP tiers.
Usage
llama.cpp / llama-server
llama-server \
--model Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-BUDGET.gguf \
--mmproj mmproj-BF16.gguf \
--ctx-size 131072 \
--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 Blackwell support.
- No
--spec-type draft-mtp— these files have no MTP head. - For vision input, add
--mmproj mmproj-BF16.gguf(from the MTP repo).
Attribution & provenance
This is a derivative work built entirely from existing Apache-2.0 artifacts. Nothing here was trained or fine-tuned. This work covers the numeric conversion only and makes no ownership claims over the model. Credit belongs to:
- Alibaba / Qwen team for the base model family (Apache-2.0): the 27B dense hybrid architecture this tune builds on.
- DavidAU for the tune itself, Qwen3.8-27B-Ultimate-Heretic-Uncensored (Apache-2.0), including the Turbo-Brilliance-50x-AGI-NEO system these files are named for.
- This repo's author for the NVFP4 conversion, the tier ladder, and stripping the MTP head.
Source provenance: stripped from revision 89cbc521c61a5043a15f6154fab856daad818edc of the upstream tune (November 16 upstream refresh, Heretic 2.1 ARA), via the same 448-tensor NVFP4 backbone as the MTP ladder.
Changelog
2026-11-20: initial release, stripped from the revised-source ladder withstrip_mtp.py(plus a one-line fix in the script to also trim the trailing entry of per-block metadata arrays such asqwen35.attention.recurrent_layers, without which llama.cpp rejects the files). BUDGET task figures were measured on the byte-identical-tensor fixed-script build; the shipped file is its deterministic rebuild (metadata-only SHA delta, tensor bytes identical, backbone re-verified 36/36 against all nine staged files).
Repository contents
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-BUDGET.gguf(14.72 GB)Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-STARVED.gguf(14.59 GB)
SHA-256
6b04d202f1583a4384ae3c610699e624827685a18ff280fac9e4779b063adb12 Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-BUDGET.gguf
7b79c749b95b9b91f6a77c6c313c92325bfc1d35d29348e3bb053faec20310dc Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-STARVED.gguf
See also
The full MTP family (ORIG + seven tiers VERY-LOW through VERY-HIGH, all with the speculative head baked in, plus mmproj-BF16.gguf) lives in esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-GGUF.
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) and DavidAU 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 or DavidAU.
Card written with AI assistance.