Modelpublic

magiccodingman/Qwen3-4B-Instruct-2507-Unsloth-MagicQuant-Hybrid-GGUF

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

8likes364downloads

Model Card

MagicQuant GGUF Hybrids - Qwen3 4B Instruct 2507

(DEPRECIATED - Part of MagicQuant v1.0 which had significant flaws. Please utilize v2.0 which is production ready)

MagicQuant is an automated quantization, benchmarking, and evolutionary hybrid-GGUF search system for LLMs.

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-K-B16-QO-Q6K-EUD-Q8_0	3.98	373.48	0.0533%
mxfp4_moe-O-Q5K-EQKUD-Q6K	3.03	428.37	0.1631%
mxfp4_moe-QUD-IQ4NL-KO-MXFP4-E-Q8_0	2.28	411.49	0.7356%
mxfp4_moe-K-B16-QU-IQ4NL-O-MXFP4-E-Q5K-D-Q6K	2.62	467.79	0.8322%
IQ4_NL	2.23	426.86	0.8996%
mxfp4_moe-EQUD-IQ4NL-KO-MXFP4	2.10	518.15	2.0904%

Table - PPL Columns

model_name	gen	gen_er	code	code_er	math	math_er
mxfp4_moe-K-B16-QO-Q6K-EUD-Q8_0	8.8766	0.2053	1.5463	0.0122	6.7119	0.1368
mxfp4_moe-O-Q5K-EQKUD-Q6K	8.8564	0.2036	1.5473	0.0122	6.6976	0.1358
mxfp4_moe-QUD-IQ4NL-KO-MXFP4-E-Q8_0	9.0127	0.2057	1.5546	0.0119	6.6919	0.1331
mxfp4_moe-K-B16-QU-IQ4NL-O-MXFP4-E-Q5K-D-Q6K	9.0490	0.2096	1.5535	0.0121	6.7221	0.1358
IQ4_NL	8.9948	0.2072	1.5600	0.0123	6.7484	0.1362
mxfp4_moe-EQUD-IQ4NL-KO-MXFP4	9.2104	0.2106	1.5598	0.0119	6.8261	0.1363

Table - Precision Loss Columns

model_name	loss_general	loss_code	loss_math
mxfp4_moe-K-B16-QO-Q6K-EUD-Q8_0	0.0720	0.0388	0.0492
mxfp4_moe-O-Q5K-EQKUD-Q6K	0.2994	0.0259	0.1640
mxfp4_moe-QUD-IQ4NL-KO-MXFP4-E-Q8_0	1.4601	0.4978	0.2489
mxfp4_moe-K-B16-QU-IQ4NL-O-MXFP4-E-Q5K-D-Q6K	1.8687	0.4267	0.2012
IQ4_NL	1.2586	0.8469	0.5933
mxfp4_moe-EQUD-IQ4NL-KO-MXFP4	3.6857	0.8339	1.7515

Baseline Models (Reference)

Table - File Size + TPS + Avg Precision Loss

model_name	file_size_gb	bench_tps	avg_prec_loss
BF16	7.50	254.70	0.0000%
Q8_0	3.99	362.48	0.0724%
Q6_K	3.08	397.92	0.2492%
Q5_K	2.69	385.17	0.7920%
IQ4_NL	2.23	426.86	0.8996%
Q4KM	2.33	377.19	0.9376%
MXFP4_MOE	2.00	467.13	8.2231%

Table - PPL Columns

model_name	gen	gen_er	code	code_er	math	math_er
BF16	8.8830	0.2056	1.5469	0.0122	6.7086	0.1369
Q8_0	8.8754	0.2053	1.5488	0.0123	6.7080	0.1367
Q6_K	8.8441	0.2034	1.5452	0.0121	6.6952	0.1357
Q5_K	8.9707	0.2079	1.5542	0.0123	6.7701	0.1384
IQ4_NL	8.9948	0.2072	1.5600	0.0123	6.7484	0.1362
Q4KM	8.9446	0.2051	1.5694	0.0125	6.7532	0.1371
MXFP4_MOE	9.8799	0.2282	1.6122	0.0130	7.3275	0.1494

Table - Precision Loss Columns

model_name	loss_general	loss_code	loss_math
BF16	0.0000	0.0000	0.0000
Q8_0	0.0856	0.1228	0.0089
Q6_K	0.4379	0.1099	0.1997
Q5_K	0.9873	0.4719	0.9167
IQ4_NL	1.2586	0.8469	0.5933
Q4KM	0.6935	1.4545	0.6648
MXFP4_MOE	11.2226	4.2213	9.2255

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 :)