Use the new version. This shown MXFP4 hybrid is no longer viable in comparison, nor really useable. The data collected is good for understanding and research, but it's not as good for real use.

Unsloth - Granite 4.0 H 1B 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 The Following Models!

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

MXFP4_MOE-output_mxfp4-router_gate_emb_q8

0% smaller than Q8 • 510.63 TPS (13.8% more TPS than Q8) • 0.1546% precision loss

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 this model 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!

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-output_mxfp4-router_gate_emb_f16	34.07%	11.48%
MXFP4_MOE-output_mxfp4-embd_f16	42.12%	13.14%
MXFP4_MOE-output_mxfp4-embd_q8	46.89%	29.89%
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	46.89%	35.83%
MXFP4_MOE-F16	41.76%	1.79%
MXFP4_MOE-output_f16-router_gate_emb_f16	33.7%	8.05%
MXFP4_MOE-output_q6_k-router_gate_emb_f16	34.07%	5.26%
Q6_K	58.97%	29.82%
MXFP4_MOE-output_q6_k-embd_f16	42.12%	14.56%
MXFP4_MOE-Q8	46.89%	19.97%
Q8_0	46.89%	23.78%
MXFP4_MOE-output_mxfp4-embd_q6_K	48.35%	41.16%
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	50.18%	31.66%
MXFP4_MOE-output_f16-router_gate_emb_q6_k	49.82%	26.81%
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	50.18%	23.31%
MXFP4_MOE-Q6_K	48.35%	35.45%
Q5_K_M	64.1%	23.91%
MXFP4_MOE-output_mxfp4-embd_q5_K	49.08%	42.51%
MXFP4_MOE-Q5_K	49.08%	36.22%
MXFP4_MOE-output_mxfp4-router_gate_emb_q5_K	52.01%	38.38%
MXFP4_MOE-output_mxfp4-router_gate_emb_q4_K	53.85%	36.7%
MXFP4_MOE-output_mxfp4-embd_q4_K	49.45%	49.46%
MXFP4_MOE-Q4_K	49.45%	35.73%
Q4_K_M	69.23%	40.58%
MXFP4_MOE-output_q8-embd_mxfp4	49.45%	29.51%
MXFP4_MOE	73.26%	47.8%

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	2.73	375.94	0
MXFP4_MOE-output_mxfp4-router_gate_emb_f16	1.8	419.09	0.1194
MXFP4_MOE-output_mxfp4-embd_f16	1.58	425.33	0.1425
MXFP4_MOE-output_mxfp4-embd_q8	1.45	488.29	0.1546
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	1.45	510.63	0.1546
MXFP4_MOE-F16	1.59	382.67	0.1846
MXFP4_MOE-output_f16-router_gate_emb_f16	1.81	406.22	0.1847
MXFP4_MOE-output_q6_k-router_gate_emb_f16	1.8	395.71	0.2094
Q6_K	1.12	488.04	0.2332
MXFP4_MOE-output_q6_k-embd_f16	1.58	430.66	0.2437
MXFP4_MOE-Q8	1.45	451.01	0.2805
Q8_0	1.45	465.34	0.2805
MXFP4_MOE-output_mxfp4-embd_q6_K	1.41	530.69	0.3201
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	1.36	494.96	0.36
MXFP4_MOE-output_f16-router_gate_emb_q6_k	1.37	476.72	0.4217
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	1.36	463.56	0.458
MXFP4_MOE-Q6_K	1.41	509.22	0.4761
Q5_K_M	0.98	465.81	0.4844
MXFP4_MOE-output_mxfp4-embd_q5_K	1.39	535.75	0.8425
MXFP4_MOE-Q5_K	1.39	512.11	0.8705
MXFP4_MOE-output_mxfp4-router_gate_emb_q5_K	1.31	520.24	0.9301
MXFP4_MOE-output_mxfp4-router_gate_emb_q4_K	1.26	513.92	3.6885
MXFP4_MOE-output_mxfp4-embd_q4_K	1.38	561.89	3.6998
MXFP4_MOE-Q4_K	1.38	510.26	3.9926
Q4_K_M	0.84	528.5	4.9497
MXFP4_MOE-output_q8-embd_mxfp4	1.38	486.87	12.9609
MXFP4_MOE	0.73	555.64	23.4457

Bench NGL was 35
Utilized CUDA

Table - PPL Columns

model_name	gen	gen_er	code	code_er	math	math_er
F16	9.7879	0.2271	1.7102	0.0136	7.7678	0.1714
MXFP4_MOE-output_mxfp4-router_gate_emb_f16	9.8235	0.227	1.7133	0.0136	7.7533	0.1706
MXFP4_MOE-output_mxfp4-embd_f16	9.8282	0.2272	1.7134	0.0136	7.7545	0.1706
MXFP4_MOE-output_mxfp4-embd_q8	9.8152	0.2269	1.7138	0.0136	7.7658	0.171
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	9.8152	0.2269	1.7138	0.0136	7.7658	0.171
MXFP4_MOE-F16	9.7995	0.2272	1.7115	0.0136	7.7957	0.1722
MXFP4_MOE-output_f16-router_gate_emb_f16	9.79	0.2269	1.7116	0.0136	7.8028	0.1724
MXFP4_MOE-output_q6_k-router_gate_emb_f16	9.8028	0.2272	1.712	0.0136	7.7966	0.1723
Q6_K	9.873	0.2293	1.7119	0.0135	7.6382	0.1661
MXFP4_MOE-output_q6_k-embd_f16	9.8146	0.2276	1.7115	0.0136	7.7975	0.1723
MXFP4_MOE-Q8	9.8094	0.2277	1.7124	0.0136	7.8061	0.1725
Q8_0	9.8094	0.2277	1.7124	0.0136	7.8061	0.1725
MXFP4_MOE-output_mxfp4-embd_q6_K	9.8638	0.2283	1.7142	0.0136	7.764	0.1707
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	9.8665	0.2284	1.7143	0.0136	7.7707	0.1709
MXFP4_MOE-output_f16-router_gate_emb_q6_k	9.8528	0.2289	1.7127	0.0136	7.8032	0.1721
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	9.8545	0.2289	1.7133	0.0136	7.8076	0.1723
MXFP4_MOE-Q6_K	9.8548	0.2289	1.7129	0.0136	7.8134	0.1724
Q5_K_M	9.9587	0.2315	1.7208	0.0137	7.697	0.168
MXFP4_MOE-output_mxfp4-embd_q5_K	9.9347	0.2296	1.716	0.0136	7.8213	0.1721
MXFP4_MOE-Q5_K	9.918	0.2299	1.715	0.0136	7.8456	0.1731
MXFP4_MOE-output_mxfp4-router_gate_emb_q5_K	9.9328	0.2294	1.7171	0.0136	7.8382	0.1726
MXFP4_MOE-output_mxfp4-router_gate_emb_q4_K	10.3958	0.2431	1.7309	0.0139	8.0509	0.1788
MXFP4_MOE-output_mxfp4-embd_q4_K	10.3658	0.2422	1.7291	0.0139	8.0855	0.1807
MXFP4_MOE-Q4_K	10.3755	0.2436	1.7277	0.0139	8.1524	0.1832
Q4_K_M	10.3225	0.2426	1.7493	0.0143	8.3194	0.187
MXFP4_MOE-output_q8-embd_mxfp4	11.8909	0.2812	1.7662	0.0143	8.8648	0.1974
MXFP4_MOE	13.7703	0.3307	1.8673	0.0163	9.3574	0.2049

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-output_mxfp4-router_gate_emb_f16	0.3637	0.1813	-0.1867
MXFP4_MOE-output_mxfp4-embd_f16	0.4117	0.1871	-0.1712
MXFP4_MOE-output_mxfp4-embd_q8	0.2789	0.2105	-0.0257
MXFP4_MOE-output_mxfp4-router_gate_emb_q8	0.2789	0.2105	-0.0257
MXFP4_MOE-F16	0.1185	0.076	0.3592
MXFP4_MOE-output_f16-router_gate_emb_f16	0.0215	0.0819	0.4506
MXFP4_MOE-output_q6_k-router_gate_emb_f16	0.1522	0.1053	0.3708
Q6_K	0.8694	0.0994	-1.6684
MXFP4_MOE-output_q6_k-embd_f16	0.2728	0.076	0.3823
MXFP4_MOE-Q8	0.2197	0.1286	0.4931
Q8_0	0.2197	0.1286	0.4931
MXFP4_MOE-output_mxfp4-embd_q6_K	0.7754	0.2339	-0.0489
MXFP4_MOE-output_mxfp4-router_gate_emb_q6_K	0.803	0.2397	0.0373
MXFP4_MOE-output_f16-router_gate_emb_q6_k	0.6631	0.1462	0.4557
MXFP4_MOE-output_q6_K-router_gate_emb_q6_K	0.6804	0.1813	0.5124
MXFP4_MOE-Q6_K	0.6835	0.1579	0.587
Q5_K_M	1.745	0.6198	-0.9115
MXFP4_MOE-output_mxfp4-embd_q5_K	1.4998	0.3391	0.6887
MXFP4_MOE-Q5_K	1.3292	0.2807	1.0016
MXFP4_MOE-output_mxfp4-router_gate_emb_q5_K	1.4804	0.4035	0.9063
MXFP4_MOE-output_mxfp4-router_gate_emb_q4_K	6.2107	1.2104	3.6445
MXFP4_MOE-output_mxfp4-embd_q4_K	5.9042	1.1051	4.09
MXFP4_MOE-Q4_K	6.0033	1.0233	4.9512
Q4_K_M	5.4618	2.2863	7.1011
MXFP4_MOE-output_q8-embd_mxfp4	21.4857	3.2745	14.1224
MXFP4_MOE	40.687	9.1861	20.464

loss_general = precision_loss_general_pct
loss_code = precision_loss_code_pct
loss_math = precision_loss_math_pct

magiccodingman/Granite-4.0-H-1B-Unsloth-MXFP4-Hybrid-GGUF

DEPRECIATED!