Nvidia-Qwen3.6-27B-NVFP4-A.gguf: The FFN and attention layers are NVFP4 quantized. The MTP draft layer is preserved at BF16. This was generated from BF16-Attn using a modified llama-quantize to re-quantize the BF16-upcast attention layers to NVFP4.
Nvidia-Qwen3.6-27B-NVFP4-Q8.gguf: The NVFP4 FFN layers are preserved, while the FP8 attention layers are converted to Q8. This was generated using convert_hf_to_gguf.py.
Nvidia-Qwen3.6-27B-NVFP4-BF16-Attn.gguf: The NVFP4 FFN layers are preserved, while the FP8 attention layers are upcast to BF16. This is the default conversion for BF16 because GGUF files do not support FP8. This was generated using convert_hf_to_gguf.py.

Quantizations provided

File	Quantization	Size
Nvidia-Qwen3.6-27B-NVFP4-A.gguf	NVFP4 FFN and attention	17.9 GB
Nvidia-Qwen3.6-27B-NVFP4-Q8.gguf	NVFP4 FFN, Q8 attention	21.5 GB
Nvidia-Qwen3.6-27B-NVFP4-BF16-Attn.gguf	NVFP4 FFN, BF16 attention	28.2 GB

Perplexity test

I tested perplexity using llama-perplexity and Salesforce's wikitext-2-raw-v1.

File	Ctx	PPL
Nvidia-Qwen3.6-27B-NVFP4-A.gguf	512	7.6619 ± 0.05299
Nvidia-Qwen3.6-27B-NVFP4-Q8.gguf	512	7.4775 ± 0.05151
Nvidia-Qwen3.6-27B-NVFP4-BF16-Attn.gguf	512	7.4814 ± 0.05157

Evaluation

The following models were evaluated for a fair comparison of capability, size and speed.

Model	Quantization	Size	Reason
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q4_K_XL	17.9 GB	Closest non-NVFP4 in size to NVFP4.
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q6_K_XL	26 GB	Closest non-NVFP4 in size to BF16-Attn.
unsloth/Qwen3.6-27B-NVFP4	NVFP4¹	25.4 GB	Alternative NVFP4 quant.

1: unsloth/Qwen3.6-27B-NVFP4 does not provide a GGUF. I used llama.cpp's conversion which passes through Unsloth's NVFP4 tensors.

	CodeFault Nvidia NVFP4-A	CodeFault Nvidia NVFP4-Q8	CodeFault Nvidia NVFP4-FB16	Unsloth NVFP4 (2026-05-31)	Unsloth UD-Q4_K_XL	Unsloth UD-Q6_K_XL
Coding
HumanEval	0.8293 ± 0.0295	0.8659 ± 0.0267	0.8354 ± 0.0290	0.8110 ± 0.0307	0.8354 ± 0.0290	0.8537 ± 0.0277
HumanEval+	0.7683 ± 0.0330	0.7683 ± 0.0330	0.7927 ± 0.0318	0.7744 ± 0.0327	0.7805 ± 0.0324	0.7805 ± 0.0324
MBPP	0.7280 ± 0.0199	0.7600 ± 0.0191	0.7540 ± 0.0193	0.7420 ± 0.0196	0.7560 ± 0.0192	0.7540 ± 0.0193
MBPP+	0.8783 ± 0.0168	0.8862 ± 0.0164	0.8836 ± 0.0165	0.8995 ± 0.0155	0.8968 ± 0.0157	0.8836 ± 0.0165
Instruction
IFEval	0.8558 ± 0.0151	0.8410 ± 0.0157	0.8410 ± 0.0157	0.8447 ± 0.0156	0.8410 ± 0.0157	0.8447 ± 0.0156
Knowledge
ARC-Challenge	0.9710 ± 0.0049	0.9701 ± 0.0050	0.9710 ± 0.0049	0.9710 ± 0.0049	0.9710 ± 0.0049	0.9710 ± 0.0049
MMLU-Pro	0.8250 ± 0.0034		0.8285 ± 0.0034	0.8293 ± 0.0033	0.8350 ± 0.0033	0.8319 ± 0.0033
STEM & Reasoning
BIG-Bench Hard	0.9149 ± 0.0032	0.9247 ± 0.0031	0.9158 ± 0.0032	0.9131 ± 0.0033	0.9260 ± 0.0030	0.9214 ± 0.0031
GPQA Diamond flexible	0.7929 ± 0.0289	0.7879 ± 0.0291	0.7828 ± 0.0294	0.8182 ± 0.0275	0.8131 ± 0.0278	0.7778 ± 0.0296
GSM8K	0.9196 ± 0.0075	0.9136 ± 0.0077	0.9136 ± 0.0077	0.9098 ± 0.0079	0.9083 ± 0.0080	0.9158 ± 0.0076
Hendrycks Math	0.3438 ± 0.0065	0.3530 ± 0.0065	0.3540 ± 0.0066	0.3662 ± 0.0066	0.3658 ± 0.0066	0.3876 ± 0.0067

These evaluations were run using lm_eval. The models were run in instruct (non-thinking) mode with the following parameters in llama-server (b9775):

ini

ctx-size = 32768
cache-type-k = q8_0
cache-type-v = q8_0
top-p = 0.8
top-k = 20
min-p = 0
presence-penalty = 1.5
spec_type = draft-mtp
spec_draft_n_max = 2
chat-template-kwargs = {"enable_thinking":false}

Benchmarks

Model	Quant	MTP n-max	Prompt Len	Output Len	Acceptance Rate	pp/s	tg/s
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-A		267	8676		2225.1	76.6
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-A	1	267	5958	0.879	1877.0	99.5
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-A	2	267	6708	0.856, 0.701	1897.2	119.1
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-A	3	267	6900	0.855, 0.699, 0.559	1950.4	128.7
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-A	4	267	5635	0.836, 0.681, 0.567, 0.483	1862.6	137.6
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-QA		267
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-QA	1	267
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-QA	2	267
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-QA	3	267
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-QA	4	267
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-BF16		267	5363		1896.4	53.4
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-BF16	1	267	5980	0.881	1643.8	74.6
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-BF16	2	267	5152	0.875, 0.732	1704.1	94.1
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-BF16	3	267	6881	0.876, 0.724, 0.595	1675.8	106.1
CodeFault/Nvidia-Qwen3.6-27B-NVFP4	NVFP4-BF16	4	267	6582	0.859, 0.692, 0.579, 0.476	1686.4	112.3
unsloth/Qwen3.6-27B-NVFP4	NVFP4		267	7347		2056.5	57.6
unsloth/Qwen3.6-27B-NVFP4	NVFP4	1	267	6826	0.843	1749.9	74.9
unsloth/Qwen3.6-27B-NVFP4	NVFP4	2	267	8142	0.851, 0.685	1794.7	90.1
unsloth/Qwen3.6-27B-NVFP4	NVFP4	3	267	7612	0.837, 0.671, 0.541	1787	96.4
unsloth/Qwen3.6-27B-NVFP4	NVFP4	4	267	7621	0.817, 0.620, 0.485, 0.400	1772.4	96.5
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q4_K_XL		267	7826		1535.4	69.8
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q4_K_XL	1	267	8398	0.879	1381.1	107.7
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q4_K_XL	2	267	7363	0.850, 0.692	1276.6	122.1
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q4_K_XL	3	267	8146	0.852, 0.681, 0.552	1286.9	123
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q4_K_XL	4	267	8269	0.830, 0.647, 0.529, 0.439	923.2	120.8
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q6_K_XL		267	7180		1257.3	53.7
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q6_K_XL	1	267	5868	0.876	1249.1	84.8
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q6_K_XL	2	267	6104	0.864, 0.701	1232.8	102.2
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q6_K_XL	3	267	5000	0.847, 0.688, 0.563	1228.7	109
unsloth/Qwen3.6-27B-MTP-GGUF	UD-Q6_K_XL	4	267	7060	0.852, 0.703, 0.577, 0.474	1052.5	116.8

These benchmarks were run on an RTX 5090 (limited to 480 W) using llama-cli (b9775) with the CUDA driver and a prompt to generate an Ansible playbook.

Serving with llama.cpp

It has a max context size of 262,114. This can be served using:

bash

llama-server \
-hf CodeFault/Nvidia-Qwen3.6-27B-NVFP4-GGUF \
-hff Nvidia-Qwen3.6-27B-NVFP4-A.gguf \
--temp 0.6 \
--top-p 0.95 \
--top-k 20 \
--repeat-penalty 1.1 \
--spec-type draft-mtp \
--spec-draft-n-max 3