MiniMax H3 FL2VA mixed-IQ1 QF2 quality report

Verdict

The original uniform minimax_h3_fl2va_pruned-IQ1_S.gguf and minimax_h3_fl2va_pruned-IQ1_M.gguf files are deprecated: perceptual FAIL. They load and produce valid audio/video containers, but neither produced a recognizable fox or snow in any frame of the controlled 22-frame test.

The QF2 mixed-IQ1 files supersede all earlier QF working candidates:

QF2 IQ1_M: PASS for the four-step website profile, including an independent motorcycle holdout. Its eight-step fox video passes visually but its audio fails relative to the matching Q8 control.
QF2 IQ1_S: PASS for the four-step website profile, including an independent motorcycle holdout. Its eight-step fox video passes visually but its audio fails relative to Q8.

These profile-specific verdicts matter: neither QF2 file is represented as having passed eight-step audiovisual generation.

Controlled quick comparison

All rows use the same fox prompt, seed 11, CPU RNG, 320x192, 22 frames at 24 FPS, 4 steps, CFG 1.0, Qwen3-VL encoder, video VAE, audio VAE, and pinned stable-diffusion.cpp runtime.

PSNR and SSIM compare decoded frames with the Q8 output at identical settings. The remaining measurements are diagnostics for this test, not universal quality scores. LPIPS was not collected and is not claimed.

Denoiser	PSNR vs Q8	SSIM vs Q8	Luma SD	Mean abs. Laplacian	Phase-grid 16	Adjacent MAD	Audio clipped	Peak dBFS	RMS dBFS	Human gate
Q8_0 control	reference	reference	41.371961	4.104783	0.0180393	3.101225	0.00%	-0.1148	-15.9386	PASS: fox and snow clearly present
Original uniform IQ1_S	11.680717 dB	0.502004	25.35	26.55	0.180	not recorded	4.50%	not recorded	not recorded	FAIL: no fox or snow in 22/22 frames
Original uniform IQ1_M	15.176109 dB	0.448978	20.56	54.10	0.403	not recorded	3.60%	not recorded	not recorded	FAIL: no fox or snow in 22/22 frames
QF2 IQ1_M	24.800133 dB	0.876201	45.441886	3.198496	0.01557654	2.260789	0.00841864%	+0.7378	-14.09494	PASS: coherent fox/snow; no grid or moire
QF2 IQ1_S	27.801705 dB	0.895095	44.8127344	3.1386273	0.0148840364	1.78388595	0.00336746%	+0.27664	-14.37595	LIMITED at low resolution: fox shape smeared, but no grid

Both QF2 variants recover sharply from the patterned-noise failure of the original uniform IQ1 files. The 640x384 website tests below are the deciding human gate for deployed use.

Website, full-profile, and holdout results

Variant and test	Settings	Visual/motion evidence	Audio evidence	Artifact SHA-256	Verdict
QF2 IQ1_M, fox website	640x384, 39 frames, 4 steps	Clear fox	RMS -15.6642 dBFS	`e42ee46827018f86e86b67238e1806a6e9d20500a90e5e03710d3cee4fa02253`	PASS; clipped fraction 0
QF2 IQ1_S, fox website	640x384, 39 frames, 4 steps	Clear walking fox; adjacent MAD 4.089718	RMS -16.70895 dBFS	`23214d5008dc58edc036d0f89b9e4a1b442be5d4265e2218d03c4416095db24a`	PASS; clipped fraction 0
QF2 IQ1_M, fox full	640x384, 39 frames, 8 steps	Excellent articulated walking fox; MAD 6.809 vs Q8 5.984	RMS -70.98 dBFS vs Q8 -50.40 dBFS	`689468e64fddec0ab0c052645e8a8a53c749a44d2a73c9d240ce5eb077ad22de`	VISUAL PASS / AUDIO FAIL
QF2 IQ1_S, fox full	640x384, 39 frames, 8 steps	Excellent articulated walking fox; MAD 6.23523; PSNR 16.112583 dB; SSIM 0.754325	RMS -73.11065 dBFS vs Q8 -50.4007 dBFS	`b8bc469b9b2f521a844b431afa750dcb5a89b9e7b88bf59123adabb0fda606af`	VISUAL PASS / AUDIO FAIL
QF2 IQ1_M, motorcycle holdout	Website profile, independent prompt, seed 123	Human PASS; PSNR 22.117554 dB; SSIM 0.885048	QF2/Q8 RMS -4.226/-4.380 dBFS; clipped 2.853%/2.530%; spectral correlation 0.8285/0.8136	`0b28a8c0a19ef8090fdc580a2c0f672ed1a341128ab0d2957f56b104a680c099`	PASS with baseline clipping disclosed
QF2 IQ1_S, motorcycle holdout	Website profile, independent prompt, seed 123	Human PASS: stable red motorcycle and rider on a coastal road at sunset; PSNR 21.096775 dB; SSIM 0.862191	QF2/Q8 RMS -4.43595/-4.37951 dBFS; clipped 2.0843%/2.5304%; spectral correlation 0.84383/0.83414; time correlation 0.6931/0.7082	`93ddcf76e9d98c78b0a51dd059a43d6bc478f451f06176ef8cda00a5a0a144f2`	A/V PASS with Q8-matched clipping disclosed

The motorcycle clipping is not hidden: Q8 also clips on that prompt. QF2 IQ1_M is 0.323 percentage points higher than Q8, while QF2 IQ1_S is 0.446 percentage points lower. This is baseline-associated behavior, not proof of clipping-free audio.

The fox prompt was included in calibration, so it is not an independent holdout. The motorcycle prompt and seed 123 were not used for calibration and provide the independent acceptance check.

Evidence separation

The published 22-frame outputs and contact sheets below document the deprecated uniform-file failure:

The public five-frame legacy files are structural smoke assets only and must not be cited as perceptual-quality evidence:

QF2 evidence is identified above by exact MP4 SHA-256 and linked from assets/qf2. The matching -contact.png files provide quick visual review. The independent Q8 holdout control is also published.

QF2 importance-guided tensor policy

QF2 is requantized from minimax_h3_fl2va_pruned-Q8_0.gguf in unsloth/MiniMax-H3-GGUF revision 9ee8213df85a2fcec53dec8c651a0fb1e821674a, using unslothai/stable-diffusion.cpp revision 13b9d92b5e9a1563536c9c980e700470f9ab6702.

The refreshed H3 activation importance matrix showed a steep rise through later fc1 blocks. First-match tensor rules therefore implement this final policy:

condition_proj.weight: BF16
all one-dimensional norm/gain tensors: retained at source precision by the patched converter
both token_refiner blocks: Q8_0 for eligible matrices
main transformer blocks 46–49: Q8_0 for every eligible matrix
main blocks 30–45, mlp.fc1: Q4_K
every main block, attention qkv_proj/out_proj and mlp.fc2: Q4_K, except blocks 46–49 already caught by the higher-priority Q8 rule
main blocks 0–29, mlp.fc1: IQ1_S or IQ1_M according to the output file
audio/video patch-projection weights: Q8_0 inherited from the Q8 source because their shapes are incompatible with IQ1/Q4_K block conversion
biases, final/output layers, and other incompatible shapes: retained at source precision by converter safeguards

Only 30 fc1 tensors remain IQ1. The final name still uses UD-IQ1_*-QF because IQ1 is the lowest precision and covers 4.624B parameters; it does not mean every tensor is one bit.

Final artifact record

Artifact	Bytes	SHA-256	Website profile	Full eight-step profile
`minimax_h3_fl2va_pruned-UD-IQ1_M-QF.gguf`	10,985,133,600	`c7993a8a0eb202f63f30cb410817fc0a504f3c844ea4e8dbb38599d29b0d4bdc`	PASS, including independent holdout	VISUAL PASS / AUDIO FAIL
`minimax_h3_fl2va_pruned-UD-IQ1_S-QF.gguf`	10,876,753,440	`b12244f203e257f10c9cec56b8062aa65d627d9250c433ef633a244d610b292c`	PASS, including independent holdout	VISUAL PASS / AUDIO FAIL

Calibration matrix: 18,693,323 bytes; SHA-256 6ce80c23f85d50cc170b1bfafa681811e5ef26590757f1c31540485b502c67a2.

Both actual QF2 GGUFs contain 532 tensors and 20,111,438,744 parameters. Their inventories are identical except for the IQ1 subtype:

Storage type	Tensors	Parameters	Policy source
BF16	212	28,113,664	protected one-dimensional tensors plus explicit BF16 `condition_proj.weight`
F16	102	43,642,368	source precision / converter safeguards
F32	8	707,224	source precision / converter safeguards
Q8_0	26	2,312,798,208	final four blocks, token refiner, and inherited audio/video patch weights
Q4_K	154	13,101,957,120	attention/out/fc2 plus importance-promoted later fc1 matrices
IQ1_S or IQ1_M	30	4,624,220,160	blocks 0–29 fc1 matrices

The audio/video patch weights are Q8_0 because the conversion input is Q8_0 and their shapes do not permit IQ1/Q4_K conversion. condition_proj.weight is separately forced to BF16.

Superseded internal QF1 evidence

These working files are not final artifacts and must not appear in the publication file table:

QF1 IQ1_M: 10,137,359,904 bytes, SHA-256 f45ebe41dedbf556a2c0b2a7885450458d26451f38a3e011a0f9572d6fb919e3; quick fox PSNR 21.790194 dB and SSIM 0.845287.
QF1 IQ1_S: 9,971,176,992 bytes, SHA-256 bca5684bb16f551a74761f5a61c9058d4986f5f0ce914535686a45850c840c12; superseded before final controlled acceptance.

QF2 replaces QF1 with importance-guided promotion of later fc1 matrices and is the only policy represented by the final filenames.

Publication gate

Both final QF2 files passed the four-step website profile and the independent motorcycle holdout. Their eight-step audio failures remain disclosed and are outside the validated publication profile.