Unsloth - Qwen3 VL 8B Thinking MXFP4 Hybrid GGUF

Dense model utilizing MXFP4_MOE with hybrid weights on a dense model. Achieving interesting results that show smaller file size, more TPS, and near lossless precision.

Use one of the 3 found magic models!

Stats compared against the standard Q8_0 (precision loss still compared to F16)

MXFP4_MOE-Q8

0% smaller than Q8 • 235.73 TPS • 0.0778% precision loss

(TLDR: I'm a scientist or something like that..)

MXFP4_MOE-output_q6_K-router_gate_emb_q6_K

10.1% smaller than Q8 • 269.24 TPS • 0.175% precision loss

(TLDR: The perfect balance)

This repository contains a set of hybrid MXFP4 quantized GGUF models designed to explore a surprising discovery:

A carefully targeted combination of MXFP4 + high-precision embeddings/output weights can deliver near-Q8 accuracy with Q4–Q6 level throughput and smaller file sizes than Q8.

Unlike pure MXFP4, which heavily degrades dense models. This hybrid method selectively protects tensors that matter most for semantic stability, while allowing MXFP4 to accelerate everything else.

This is experimental. And should be treated as such. I am more than encouraging people to use these models and leave feedback! Though precision loss seemed near lossless, did the hybrid models act strange in certain situations? Worse or better on some topics compared to the original model? Did it do better/worse overall on everything? I'd love to hear back from others!

The Magic Models - Use One Of These 3 Models!

Each of these models achieved:

File size reduction compared to the Q8_0

Better precision loss scores than the pure Q6_K

Achieving noticeably better TPS than a Q4_K_M

I have personally deemed these in the category of "Q7.5" quantization.

The following are the special models to note from what was created. Each of the 3 models shown below are being compared to the Q8 model.

MXFP4_MOE-Q8

(0% smaller than Q8 • 235.73 TPS • 0.0778% precision loss)

The TPS and precision loss difference between this model and the standard Q8_0 is negligable and potentially within margin of error. But safe to use, but honestly not much different from a normal Q8_0 model.

The following was the conversion script:

llama-quantize \
  --tensor-type token_embd.weight=Q8_0 \
  --tensor-type output.weight=Q8_0 \
  "Path_To_F16_GGUF.gguf" \
  "Path_To_GGUF.gguf" \
  mxfp4_moe

MXFP4_MOE-output_q6_K-router_gate_emb_q6_K

(10.1% smaller than Q8 • 269.24 TPS • 0.175% precision loss)

This version has very minor precision loss and has a fantastic TPS.

The following was the conversion script:

llama-quantize \
  --tensor-type token_embd.weight=Q6_K \
  --tensor-type output.weight=Q6_K \
  --tensor-type 'router.*'=Q6_K \
  --tensor-type 'gate.*'=Q6_K \
  "Path_To_F16_GGUF.gguf" \
  "Path_To_GGUF.gguf" \
  mxfp4_moe

MXFP4_MOE Hybrid Naming Scheme & Synopsis

Multiple different combinations of converted models were created. The results were interesting to say the least. The following table will explain my naming scheme to what was done to the model to create it.

Suffix Example	Meaning
`MXFP4_MOE`	Pure MXFP4 pipeline
`MXFP4_MOE-Q8`	Embedding/output in Q8_0
`MXFP4_MOE-F16`	Embedding/output in F16
`output_mxfp4-embd_q8`	Output → MXFP4, Embedding → Q8
`output_mxfp4-router_gate_emb_q5_K`	Output → MXFP4, Emb/Router/Gate → Q5_K
`MXFP4_MOE-Q6_K`	Both embedding + output in Q6_K
`Q8_0`, `Q6_K`, `Q4_K_M`	Pure model-wide quantizations

The results achieved were interesting to say the least. It was a brute force game of mass creating models with hybrid methods to find combinations that didn't cause too much noise and paired well with MXFP4.

This repo showcases the converted models, whether good or bad that was created. But, I have been testing other models in different combinations as well. The winning hybrid combinations shown in this repo DOES NOT always equate to the same results on different models.

Some models do better or worse with different kinds of combinations. It depends if it's dense, MOE, and much more. Many times the results surprise me. Many models no matter the combination will not play nice with MXFP4. At least with the methods shown here.

Benchmark Methodology

All models were tested with a unified automated harness using llama.cpp tools.

Included tests:

Throughput:
llama-bench with descending GPU offload (-ngl 35 → 0) and automatic OOM retry.
Highest successful TPS is recorded.
Perplexity:
Three domains: general, code, math.
Each uses an auto-generated corpus of ~32k tokens.
Perplexity is computed with llama-perplexity at 2048-token context.
Same GPU retry logic as above.
Precision loss:
Each model is compared to its family F16 baseline.
Precision-loss % is computed for all PPL domains, plus an averaged score.
Models are ranked by this metric.

Table - Overview of Results

Comparing to F16.

model_name	size_reduction	tps_change
MXFP4_MOE-F16	36.24%	6.88%
MXFP4_MOE-Q8	46.85%	46.9%
Q8_0	46.85%	46.27%
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	52.23%	67.78%
MXFP4_MOE-output_q8-embd_mxfp4	48.89%	51.19%
MXFP4_MOE-Q6_K	49.61%	66.83%
Q6_K	58.98%	47.07%
MXFP4_MOE-Q5_K	51.05%	52.34%
Q5_K_M	64.29%	55.11%
MXFP4_MOE-Q4_K	52.49%	76.66%
Q4_K_M	69.33%	66.13%
MXFP4_MOE-output_mxfp4-embd_q6_K	51.77%	89.2%
MXFP4_MOE-output_mxfp4-embd_q5_K	52.23%	66.38%
MXFP4_MOE-output_mxfp4-router_gate_emb_q5_K	56.42%	82.72%
MXFP4_MOE-output_mxfp4-embd_q8	50.85%	81.53%
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	50.85%	80.79%
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	54.39%	89.95%
MXFP4_MOE-output_mxfp4-embd_q4_K	52.75%	81.3%
MXFP4_MOE-output_mxfp4-router_gate_emb_q4_K	58.26%	89.38%
MXFP4_MOE	73.39%	107.35%

All percentages compared against the selected family F16 baseline.

Table - File Size + TPS + Avg Precision Loss

model_name	file_size_gb	bench_tps	avg_prec_loss
F16	15.26	160.47	0
MXFP4_MOE-F16	9.73	171.51	0.0625
MXFP4_MOE-Q8	8.11	235.73	0.0778
Q8_0	8.11	234.72	0.082
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	7.29	269.24	0.175
MXFP4_MOE-output_q8-embd_mxfp4	7.8	242.62	0.1971
MXFP4_MOE-Q6_K	7.69	267.72	0.2022
Q6_K	6.26	236.01	0.2571
MXFP4_MOE-Q5_K	7.47	244.46	0.3489
Q5_K_M	5.45	248.9	0.5776
MXFP4_MOE-Q4_K	7.25	283.49	1.0938
Q4_K_M	4.68	266.59	1.2473
MXFP4_MOE-output_mxfp4-embd_q6_K	7.36	303.61	3.0781
MXFP4_MOE-output_mxfp4-embd_q5_K	7.29	266.99	3.0899
MXFP4_MOE-output_mxfp4-router_gate_emb_q5_K	6.65	293.21	3.0959
MXFP4_MOE-output_mxfp4-embd_q8	7.5	291.3	3.1008
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	7.5	290.12	3.1008
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	6.96	304.81	3.107
MXFP4_MOE-output_mxfp4-embd_q4_K	7.21	290.93	3.12
MXFP4_MOE-output_mxfp4-router_gate_emb_q4_K	6.37	303.9	3.2112
MXFP4_MOE	4.06	332.74	9.1432

Bench NGL was 35
Utilized CUDA

Table - PPL Columns

model_name	gen	gen_er	code	code_er	math	math_er
F16	7.9733	0.172	1.4317	0.0092	5.9985	0.1097
MXFP4_MOE-F16	7.9857	0.1724	1.4318	0.0092	6	0.1097
MXFP4_MOE-Q8	7.9846	0.1723	1.4322	0.0093	6.0019	0.1097
Q8_0	7.9848	0.1723	1.4323	0.0093	6.0021	0.1097
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	8.0038	0.1729	1.4325	0.0093	6.0037	0.1097
MXFP4_MOE-output_q8-embd_mxfp4	7.9958	0.1726	1.4335	0.0093	6.0095	0.1099
MXFP4_MOE-Q6_K	8.0054	0.173	1.4324	0.0093	6.0078	0.1099
Q6_K	8.0126	0.173	1.4333	0.0093	6.0085	0.1098
MXFP4_MOE-Q5_K	8.0006	0.173	1.433	0.0092	6.0353	0.1105
Q5_K_M	8.0504	0.1739	1.4377	0.0093	6.0193	0.11
MXFP4_MOE-Q4_K	8.1239	0.1764	1.4371	0.0094	6.0594	0.1115
Q4_K_M	8.105	0.1743	1.4495	0.0095	6.0493	0.1104
MXFP4_MOE-output_mxfp4-embd_q6_K	8.2126	0.1751	1.4406	0.0092	6.3351	0.1157
MXFP4_MOE-output_mxfp4-embd_q5_K	8.2156	0.1751	1.4409	0.0092	6.3337	0.1157
MXFP4_MOE-output_mxfp4-router_gate_emb_q5_K	8.2042	0.1747	1.4423	0.0092	6.3375	0.1158
MXFP4_MOE-output_mxfp4-embd_q8	8.2169	0.1752	1.4409	0.0092	6.3347	0.1157
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	8.2169	0.1752	1.4409	0.0092	6.3347	0.1157
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	8.2222	0.1753	1.441	0.0092	6.3314	0.1156
MXFP4_MOE-output_mxfp4-embd_q4_K	8.216	0.1752	1.441	0.0092	6.3384	0.1158
MXFP4_MOE-output_mxfp4-router_gate_emb_q4_K	8.213	0.1752	1.4446	0.0093	6.342	0.116
MXFP4_MOE	8.8566	0.1912	1.4951	0.01	6.7137	0.125

gen = ppl_general
gen_er = ppl_general_error
code = ppl_code
code_er = ppl_code_error
math = ppl_math
math_er = ppl_math_error

Table - Precision Loss Columns

model_name	loss_general	loss_code	loss_math
F16	0	0	0
MXFP4_MOE-F16	0.1555	0.007	0.025
MXFP4_MOE-Q8	0.1417	0.0349	0.0567
Q8_0	0.1442	0.0419	0.06
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	0.3825	0.0559	0.0867
MXFP4_MOE-output_q8-embd_mxfp4	0.2822	0.1257	0.1834
MXFP4_MOE-Q6_K	0.4026	0.0489	0.155
Q6_K	0.4929	0.1118	0.1667
MXFP4_MOE-Q5_K	0.3424	0.0908	0.6135
Q5_K_M	0.967	0.4191	0.3468
MXFP4_MOE-Q4_K	1.8888	0.3772	1.0153
Q4_K_M	1.6518	1.2433	0.8469
MXFP4_MOE-output_mxfp4-embd_q6_K	3.0013	0.6216	5.6114
MXFP4_MOE-output_mxfp4-embd_q5_K	3.0389	0.6426	5.5881
MXFP4_MOE-output_mxfp4-router_gate_emb_q5_K	2.8959	0.7404	5.6514
MXFP4_MOE-output_mxfp4-embd_q8	3.0552	0.6426	5.6047
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	3.0552	0.6426	5.6047
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	3.1217	0.6496	5.5497
MXFP4_MOE-output_mxfp4-embd_q4_K	3.0439	0.6496	5.6664
MXFP4_MOE-output_mxfp4-router_gate_emb_q4_K	3.0063	0.901	5.7264
MXFP4_MOE	11.0782	4.4283	11.923