Datasheet for local AI49 models98 GPUsData read 2026-09-25
No adsNo tracking
Check my GPU →

Can the RTX 5090 32 GB run MiniMax H3 Pruned?

Video model · pruned 33B32 GB GDDR7 · 512-bit · 1792 GB/s · BlackwellData 2026-09-25
Runs well
Yes — and comfortably.

Download FP8 (21.0 GB). With the model's working memory it needs about 26.8 GB, leaving 5.2 GB spare on 32 GB. Quality: practically identical to the original.

Best fileFP8
File size21.0 GB
VRAM needed~26.8 GB
System RAM64 GB+
Memory map · RTX 5090 32 GB26.8 GB / 32 GB
016 GB32 GB
Weights FP8 · 21.0 GBWorking memory · 5.0 GBReserve · 0.8 GBFree · 5.2 GB
calculated from real file sizes plus working memory. Real-world results people have reported are listed further down. How this works.

01What to download

PartFileFolderSize
Modelminimax_h3_fl2va_pruned_fp8_scaled.safetensors
FP8 · Comfy-Org/MiniMax-H3
models/diffusion_models21.0 GBDownload →
Text encoderqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
Qwen3-VL 32B NVFP4 (AWQ) · Comfy-Org/MiniMax-H3
models/text_encoders15.7 GBDownload →
VAEminimax_h3_video_vae_fp16.safetensors
MiniMax H3 video VAE FP16 · Comfy-Org/MiniMax-H3
models/vae5.2 GBDownload →
Also neededminimax_h3_audio_vae_fp32.safetensors
MiniMax H3 audio VAE FP32 · Comfy-Org/MiniMax-H3
models/vae0.6 GBDownload →
Total download · keep about the same free on disk42.5 GB

Alternatives: Qwen3-VL 32B INT8 (convrot) (27.1 GB); Qwen3-VL 32B BF16 (51.5 GB). The 16-bit text encoder is a little more faithful; the smaller one is chosen here because it loads faster and needs less RAM. Sizes read from Hugging Face (2026-09-25). “Download” links start the file directly; the file name links are the same files the official ComfyUI workflows use.

System RAM: 64 GB or more. ComfyUI keeps the model file, the text encoder and the VAE in system RAM and moves them to the GPU as needed. With this set of files that is about 42.5 GB, plus roughly 6 GB for Windows, ComfyUI and a browser: 48.5 GB in total. With less RAM it still runs, but Windows starts swapping to disk and loading gets very slow. calculated

02Every MiniMax H3 Pruned file on 32 GB

FileSizeNeededOn this cardQualityDownload
16-bit
SAFETENSORS · Comfy-Org
40.2 GB46.0 GB14.0 GB too bigthe original weightsHugging Face →
FP8 ←
SAFETENSORS · Comfy-Org
21.0 GB26.8 GBfits · 5.2 GB sparepractically identical to the originalHugging Face →
Q8_0
GGUF · Abiray
21.6 GB27.4 GBfits · 4.6 GB sparepractically identical to the originalHugging Face →
INT8
SAFETENSORS · Comfy-Org
21.0 GB26.8 GBfits · 5.2 GB sparepractically identical to the originalHugging Face →
Q6_K
GGUF · Abiray
16.7 GB22.5 GBfits · 9.5 GB sparevery close to the originalHugging Face →
Q5_K_M
GGUF · Abiray
14.1 GB19.9 GBfits · 12.1 GB spareclose; small differences in fine detailHugging Face →
Q4_K_M
GGUF · Abiray
11.6 GB17.4 GBfits · 14.6 GB sparegood; some loss in fine detail and textHugging Face →
Q3_K_M
GGUF · Abiray
8.9 GB14.7 GBfits · 17.3 GB sparenoticeable loss of detailHugging Face →

Sizes from the Hugging Face file listing, read 2026-09-25. “Needed” = file + 5 GB working memory for this model + 0.8 GB kept free for the system.

03The text encoder

Qwen3-VL 32B: 51.5 GB as 16-bit, 27.1 GB as INT8, 15.7 GB as NVFP4. ComfyUI encodes the prompt first and can push the encoder out of VRAM before sampling, so it does not have to fit together with the model. On 32 GB the INT8 encoder fits on its own, so prompt encoding stays fast.

04About this GPU

The RTX 5090 32 GB is a Blackwell GPU with FP8 and FP4 hardware: ComfyUI computes Comfy-Org's FP8 files natively here, and NVFP4 files (where a model offers them) are faster still.

RTX 5090 32 GB: 32 GB GDDR7 · 512-bit · 1792 GB/s · Blackwell. Everything that runs on the RTX 5090 32 GB →

05What more VRAM would change

Nothing to gain for this model: the RTX 5090 32 GB already runs a top-quality file entirely in VRAM. More memory would only help with bigger images, longer clips or several models at once.

06Measured and reported results

LabelSetupResultPeak VRAMDateSource
reportedminimax_h3_fl2va_pruned_nvfp4 · 864x480 · 10 steps
“175 s for a 864×480 ten-second clip, 26,914 MiB peak VRAM.”
ComfyUI 0.30.1, 500 W power cap; 26,914 MiB = 26.28 GiB; INT8-ConvRot = 185 s / 28,581 MiB
175 s / clip (243 frames)26.28 GB2026-08-04ai-muninn.com →

Reported results are other people's numbers, copied as published, with a link. Settings, drivers and ComfyUI versions differ, so compare them with care. Send yours.

07Questions

How much VRAM does MiniMax H3 Pruned need?

Around 14.7 GB with the smallest file (Q3_K_M) and 26.8 GB with an 8-bit file (FP8), counting working memory and a small system reserve. The full 16-bit file needs about 46.0 GB.

Which MiniMax H3 Pruned file should I download for the RTX 5090 32 GB?

FP8 (21.0 GB) from Comfy-Org/MiniMax-H3.

Is FP8 faster than GGUF on the RTX 5090 32 GB?

It can be. This GPU has FP8 hardware, and ComfyUI computes FP8 natively for files made for it (Comfy-Org's fp8_scaled files), or for any FP8 file with the --fast fp8_matrix_mult option. GGUF files are unpacked on the fly, which costs some speed.

How much do I need to download for MiniMax H3 Pruned on the RTX 5090 32 GB?

About 42.5 GB for the model file, text encoder and VAE listed on this page, and the same again free on disk. With a 32 GB card, 64 GB of system RAM or more is recommended.