Each release includes models optimized to outperform standard baseline quants (Q8, Q6, Q5, Q4). If a baseline GGUF exists in this repo, the evolutionary engine couldn’t beat it. If a baseline is missing, it’s because a hybrid configuration outperformed it so completely that including the baseline would've been pointless.

These hybrid GGUFs are built to be as small, fast, and low-drift as possible while preserving model capability.

To dive deeper into how MagicQuant works, see the main repo: MagicQuant on GitHub (by MagicCodingMan)

Notes:

The HuggingFace hardware compatibility where it shows the bits is usually wrong. It doesn't understand hybrid mixes, so don't trust it.
Naming scheme can be found on the MagicQuant Wiki.
(tips) Less precision loss means less brain damage. More TPS means faster! Smaller is always better right?

Precision Loss Guide

0–0.1% → God-tier, scientifically exact
0.1–1% → True near-lossless, agent-ready
1–3% → Minimal loss, great for personal use
3–5% → Borderline, but still functional
5%+ → Toys, not tools, outside MagicQuant’s scope

Learn more about precision loss here.

Table - File Size + TPS + Avg Precision Loss

model_name	file_size_gb	bench_tps	avg_prec_loss
mxfp4_moe-HQKOR-B16-U-Q5K-E-Q6K-D-Q8_0	36.31	85.41	0.0223%
Q8_0	30.25	99.66	0.1182%
Q5_K	20.23	123.94	0.2558%
mxfp4_moe-H-B16-EUD-IQ4NL-R-Q6K-QKO-Q8_0	19.20	115.33	0.4621%
iq4_nl-QKOUD-IQ4NL-EH-Q8_0	16.33	145.90	0.8683%
iq4_nl-QKOUD-IQ4NL-E-MXFP4-H-Q5K	16.07	153.05	1.1878%

Table - PPL Columns

model_name	gen	gen_er	code	code_er	math	math_er
mxfp4_moe-HQKOR-B16-U-Q5K-E-Q6K-D-Q8_0	6.2842	0.1284	1.2904	0.0068	5.6809	0.1047
Q8_0	6.2952	0.1287	1.2894	0.0069	5.6903	0.1050
Q5_K	6.3057	0.1289	1.2963	0.0069	5.6818	0.1045
mxfp4_moe-H-B16-EUD-IQ4NL-R-Q6K-QKO-Q8_0	6.3141	0.1294	1.2965	0.0070	5.7085	0.1055
iq4_nl-QKOUD-IQ4NL-EH-Q8_0	6.3539	0.1294	1.3056	0.0071	5.7017	0.1040
iq4_nl-QKOUD-IQ4NL-E-MXFP4-H-Q5K	6.3772	0.1301	1.3056	0.0071	5.7351	0.1051

Table - Precision Loss Columns

model_name	loss_general	loss_code	loss_math
mxfp4_moe-HQKOR-B16-U-Q5K-E-Q6K-D-Q8_0	0.0573	0.0078	0.0018
Q8_0	0.1177	0.0698	0.1672
Q5_K	0.2847	0.4650	0.0176
mxfp4_moe-H-B16-EUD-IQ4NL-R-Q6K-QKO-Q8_0	0.4183	0.4805	0.4876
iq4_nl-QKOUD-IQ4NL-EH-Q8_0	1.0512	1.1858	0.3679
iq4_nl-QKOUD-IQ4NL-E-MXFP4-H-Q5K	1.4218	1.1858	0.9559

Baseline Models (Reference)

Table - File Size + TPS + Avg Precision Loss

model_name	file_size_gb	bench_tps	avg_prec_loss
BF16	56.90	51.02	0.0000%
Q8_0	30.25	99.66	0.1182%
Q5_K	20.23	123.94	0.2558%
Q6_K	23.37	114.97	0.2965%
IQ4_NL	16.26	138.47	1.0534%
Q4_K_M	17.28	130.97	1.3851%
MXFP4_MOE	15.15	141.87	10.2733%

Table - PPL Columns

model_name	gen	gen_er	code	code_er	math	math_er
BF16	6.2878	0.1285	1.2903	0.0069	5.6808	0.1047
Q8_0	6.2952	0.1287	1.2894	0.0069	5.6903	0.1050
Q5_K	6.3057	0.1289	1.2963	0.0069	5.6818	0.1045
Q6_K	6.3172	0.1294	1.2927	0.0069	5.6942	0.1051
IQ4_NL	6.3497	0.1293	1.3042	0.0070	5.7432	0.1057
Q4_K_M	6.4310	0.1316	1.3029	0.0070	5.7320	0.1055
MXFP4_MOE	7.1681	0.1508	1.3566	0.0080	6.3444	0.1214

Table - Precision Loss Columns

model_name	loss_general	loss_code	loss_math
BF16	0.0000	0.0000	0.0000
Q8_0	0.1177	0.0698	0.1672
Q5_K	0.2847	0.4650	0.0176
Q6_K	0.4676	0.1860	0.2359
IQ4_NL	0.9844	1.0773	1.0984
Q4_K_M	2.2774	0.9765	0.9013
MXFP4_MOE	14.0001	5.1383	11.6815

Support

I’m a solo developer working full time for myself to achieve my dream, pouring nights and weekends into open protocols and tools that I hope make the world a little better. If you chip in, you're helping me keep the lights on while I keep shipping.

Click here to see ways to support - BTC, Paypal, GitHub sponsors.

Or, just drop a like on the repo :)

magiccodingman/Qwen3-30B-A3B-Thinking-2507-unsloth-MagicQuant-Hybrid-GGUF

MagicQuant GGUF Hybrids - Qwen3 30B A3B Thinking 2507

Table - File Size + TPS + Avg Precision Loss

Table - PPL Columns

Table - Precision Loss Columns

Baseline Models (Reference)

Table - File Size + TPS + Avg Precision Loss

Table - PPL Columns

Table - Precision Loss Columns

Support

magiccodingman/
Qwen3-30B-A3B-Thinking-2507-unsloth-MagicQuant-Hybrid-GGUF