A family of seven GGUF files of DavidAU/Qwen3.8-27B-Ultimate-Heretic-Uncensored, DavidAU's Ultimate-Heretic Turbo-Brilliance-50x-AGI-NEO tune of Qwen3.8-27B: a 27B dense hybrid model (Gated DeltaNet plus Gated Attention every fourth layer, embedded MTP speculative head). The MTP head is baked into every file; no separate drafter is needed (--spec-type draft-mtp).
HF repo: esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-GGUF
The classic 256-tensor NVFP4 checkpoint this family was built from lives in esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4 with its own card.
The 400-tensor W4A16 serving checkpoint lives in esatapedico/Qwen3.8-27B-Ultimate-Heretic-Uncensored-Turbo-Brilliance-50x-AGI-NEO-NVFP4-W4A16-T400 with its own card.
My part here is only the numerics: I converted the NVFP4 checkpoint to GGUF and built a size and precision ladder for the tensors that most affect output quality and decode speed. All credit for the model itself belongs upstream (full chain below).
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.
Every tier carries the same byte-identical 448-tensor NVFP4 backbone (per-tensor SHA-256 verified) and differs only in the precision of the LM head, token embedding, and MTP draft head:
| File | Size | lm_head (output.weight) |
token_embd | MTP head (blk.64) | Backbone | SHA-256 |
|---|---|---|---|---|---|---|
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-VERY-LOW.gguf |
14,862,508,192 bytes (14.86 GB) | Q3_K |
Q2_K |
Q2_K |
NVFP4 |
37ca232fe0700dd6eeb25fe0563c971b4faeeb96ea5447e1966da91887389d96 |
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-COMPACT-LOW.gguf |
15,160,492,192 bytes (15.16 GB) | Q4_K |
Q3_K |
Q2_K |
NVFP4 |
5a3abdd28931f79a4b6d31fdfa148790fcf5b272139ee0ccd62f5bb6688f838f |
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-LOW.gguf |
15,534,805,152 bytes (15.53 GB) | Q5_0 |
IQ4_XS |
IQ4_XS |
NVFP4 |
1147fd3a4e8af30d6b777a405564c78f62dca53c67c7f9bbe093a57fc55c2aa9 |
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-MEDIUM.gguf |
16,379,093,152 bytes (16.38 GB) | Q8_0 |
Q6_K |
IQ4_XS |
NVFP4 |
b97aaf359cb4a3bfff7e0cbabbe0aaf53561a4c65265d0919156cc5cb522065d |
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-MID-HIGH.gguf |
16,912,617,632 bytes (16.91 GB) | Q8_0 |
Q8_0 |
Q8_0 |
NVFP4 |
60adecc623755577ecc4a5f2eb78a39fc9a58a90fd4fa8bf9ac04e7d56092231 |
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-HIGH.gguf |
17,571,029,120 bytes (17.57 GB) | BF16 |
Q6_K |
IQ4_XS |
NVFP4 |
a9d2a4b3f602580a830366cc2f67b68ad5f18bc85772b41b8570fab87af92b6c |
Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-VERY-HIGH.gguf |
19,694,620,832 bytes (19.69 GB) | BF16 |
BF16 |
BF16 |
NVFP4 |
d114c4fffd4e66bb675263f81f986eb4e87fd3df96e501a6461a1e322ed9d951 |
Picking a tier: MID-HIGH is the highest-precision compact option (all three head groups at Q8_0). COMPACT-LOW sits between VERY-LOW and LOW with a stronger LM head than VERY-LOW (Q4_K versus Q3_K). LOW and VERY-LOW trade some head precision for less VRAM. HIGH and VERY-HIGH restore BF16 heads where VRAM allows.
Each tier holds 1,122 tensors. The shared 448-tensor NVFP4 backbone is byte-identical across all seven tiers (per-tensor SHA-256 verified): 256 tensors quantized in the NVFP4 conversion pass plus 192 normalized to NVFP4 when the tiers were built. The 48 ssm_alpha vectors stay F32.
| File | Size | Notes | SHA-256 |
|---|---|---|---|
mmproj-BF16.gguf |
927,608,672 bytes (0.93 GB) | Vision projector, pair with any tier via --mmproj |
e4ac3e71d099c4b1d6274558fabe2edd42a0cac4f0bce1c13924bc7a7a1e9314 |
Pair any tier with mmproj-BF16.gguf via --mmproj. Vision serving verified (MID-HIGH answers image prompt, natural stop).
Perplexity on a source-diverse corpus (lower is better):
| Tier | PPL |
|---|---|
| VERY-LOW | 3.2874±0.07455 |
| COMPACT-LOW | 3.2841±0.07468 |
| LOW | 3.2751±0.07442 |
| MEDIUM | 3.2596±0.07381 |
| MID-HIGH | 3.2569±0.07359 |
| HIGH | 3.2602±0.07384 |
| VERY-HIGH | 3.2441±0.07320 |
Perplexity stays within about 1.3% of the unquantized 3.2445 figure across the whole ladder, so the head ladder costs almost nothing even at the smallest tier.
MMLU letter-logprob (thinking off) and GSM8K-200, identical flags every tier:
| Tier | MMLU (n=1000) | GSM8K (n=200) |
|---|---|---|
| VERY-LOW | 83.09% | 0.940 |
| COMPACT-LOW | 84.15% | 0.935 |
| LOW | 84.86% | 0.940 |
| MEDIUM | 84.50% | 0.940 |
| MID-HIGH | 83.47% | 0.935 |
| HIGH | 84.17% | 0.935 |
| VERY-HIGH | 82.85% | 0.935 |
MMLU sits within sampling noise across the ladder (SE about 1.2pp at n=1000) while perplexity improves monotonically with size. MID-HIGH reproduced its earlier 83.47% gate figure exactly.
llama-server \
--model Qwen3.8-27B-Ultimate-Heretic-Turbo-Brilliance-50x-AGI-NEO-NVFP4-MTP-MEDIUM.gguf \
--ctx-size 131072 \
--flash-attn on \
--spec-type draft-mtp \
--temp 0.7 --top-p 0.95 --top-k 20
draft-mtp spec path for the embedded MTP head.--mmproj mmproj-BF16.gguf (verified with MID-HIGH).Linear, ignore list lm_head, visual tower, and the full linear_attn path. The MTP head is excluded from quantization and restored to BF16 after. Sequential run, CPU only, no calibration corpus.convert_hf_to_gguf.py --outtype auto, keeping native NVFP4 tensors intact.llama-quantize --tensor-type-file over the shared backbone.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:
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.
有文件?提交种子。
提交种子