Intro

I usually test T2I models, and this one is video model, but it is also capable T2I and latest SD.Next is provide this capability so I can test it. This is 33B (69B in Total parameters) model that is not running on my mini-PC 128GB unified memory in bf16. So, for test quantized models will be used.

Hugging Face repo links

• MiniMaxAI/MiniMax-H3 498 GB

• OzzyGT/MiniMax_H3_sdnq_dynamic_8bit 108 GB

• OzzyGT/MiniMax_H3_sdnq_dynamic_4bit 64.8 GB

• OzzyGT/MiniMax_H3_sdnq_8bit_pruned 40.4 GB

• OzzyGT/MiniMax_H3_sdnq_4bit_pruned 23.7 GB

CivitAi model

• https://civitai.red/models/2821932/minimax (Video only)

Total Parameters

Total parameters compared to other large (home lab kind of large) models

• Microsoft Lens 4B - 25'103'028'451

• Hunyuan Image 2.1 17B -  26'342'852'611

• Ideogram 4  9B -  26'808'357'379

• Qwen Image 2512 20B - 28'849'460'275

• HiDream I1 Full 17B - 30'800'746'723

• FLUX.2 Dev 32B - 56'318'688'547

• MiniMax H3 33B 69'235'573'433

Resolutions

Image creation time

Work (USABLE RESULTS) in at least 512 to 1920 1:1 resolution (see the test).

Time to produce image on Intel Core Ultra 9 285H iGPU for CFG1/50STEPS

4bit version

33B time (20B pruned result)

• 128px - 4m 4s (2m29s)

• 256px - 4m 25s (2m 46s)

• 512px - 6m 23s (4m 24s)

• 704px - 9m 18s (6m 33s)

• 1024px - 14m 14s (10m 44s )

• 1280px 18m 25s (15m 49s)

• 1536px - 24m 12s (21m 30s)

• 1920px - 38m 30s (33m 56s)

8bit version

• 128px - 4m 10s (2m42s)

• 256px - 4m 29s (3m 0s)

• 512px - 6m 6s (4m 38s)

• 704px - 8m 27s (7m 7s)

• 1024px - 13m 1s (11m 45s )

• 1280px 18m 23s (17m 0s)

• 1536px -24 m 15s (22m 30s)

• 1920px - 38m 49s (37m 56s)

For reference KREA 2 Turbo (bf16) 1024px take 2m 13.57s for 8 STEP on this HW.

128x128

256x256

512x512

704x704

1024x1024

1280x1280

1536x1536

1920x1920

Image Quality

1920px

Dynamic 4bit 1920px image examples

that is ok (1024px version is https://civitai.red/images/139486776)

https://civitai.red/images/139620400 also kind of ok

https://civitai.red/images/139620409 not bad

https://civitai.red/images/139620415 biting lips, wrong, but still nice

https://civitai.red/images/139620424 most of open model doing wrong. but this one is ok

1024px

text is correct

comics also

Solar system is too complex for this model

https://civitai.red/images/139535021 Gangster duck is good

https://civitai.red/images/139535019

https://civitai.red/images/139571934 foreshortening is good.

illustration example

Manhattan bridge is low quality (Oslo opera is also both wrong and low quality)

more images in test v1/v2/v3 (see Links sections)

Links

CivitAi Posts

• https://civitai.red/posts/30366533 or https://civitai.com/posts/30366533

• https://civitai.red/posts/30357074 or https://civitai.com/posts/30357074

• https://civitai.red/posts/30339923 or https://civitai.com/posts/30339923

• https://wiki.liutyi.info/display/AI/MiniMax+H3+sdnq+dynamic+4bit

• https://wiki.liutyi.info/display/AI/MiniMax+H3+sdnq+dynamic+4bit+test+v2

• https://wiki.liutyi.info/display/AI/MiniMax+H3+sdnq+dynamic+4bit+test+v3