Can the RTX 4090 24 GB run MiniMax H3 Pruned?
Download Q6_K (16.7 GB). With the model's working memory it needs about 22.5 GB, leaving 1.5 GB spare on 24 GB. Quality: very close to the original.
Worth trying on this GPU: with an up-to-date ComfyUI, Dynamic VRAM (on by default for NVIDIA since March 2026) streams whatever does not fit from system RAM, and ComfyUI's own start-up message recommends native FP8/INT8 files over GGUF, saying they “will be faster even if they are larger than your memory”. So before settling for a heavily compressed GGUF, try the native FP8 file (21.0 GB). The file recommended above is the best one that fits entirely — the safe choice on older ComfyUI versions, AMD and Intel. Comfy blog: Dynamic VRAM →
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | MiniMax-H3-FL2VA-Pruned-Q6_K.gguf Q6_K · Abiray/MiniMax-H3-Pruned-GGUF | models/unet | 16.7 GB | Download → |
| Text encoder | qwen3vl_32b_minimax_h3_int8_convrot.safetensors Qwen3-VL 32B INT8 (convrot) · Comfy-Org/MiniMax-H3 | models/text_encoders | 27.1 GB | Download → |
| VAE | minimax_h3_video_vae_fp16.safetensors MiniMax H3 video VAE FP16 · Comfy-Org/MiniMax-H3 | models/vae | 5.2 GB | Download → |
| Also needed | minimax_h3_audio_vae_fp32.safetensors MiniMax H3 audio VAE FP32 · Comfy-Org/MiniMax-H3 | models/vae | 0.6 GB | Download → |
| Total download · keep about the same free on disk | 49.7 GB | |||
Alternatives: Qwen3-VL 32B NVFP4 (AWQ) (15.7 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. FP4 files are listed only as alternatives unless the GPU is an RTX 50 (Blackwell). 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 49.7 GB, plus roughly 6 GB for Windows, ComfyUI and a browser: 55.7 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 24 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit SAFETENSORS · Comfy-Org | 40.2 GB | 46.0 GB | 22.0 GB too big | the original weights | Hugging Face → |
| FP8 SAFETENSORS · Comfy-Org | 21.0 GB | 26.8 GB | 2.8 GB too big | practically identical to the original | Hugging Face → |
| Q8_0 GGUF · Abiray | 21.6 GB | 27.4 GB | 3.4 GB too big | practically identical to the original | Hugging Face → |
| INT8 SAFETENSORS · Comfy-Org | 21.0 GB | 26.8 GB | 2.8 GB too big | practically identical to the original | Hugging Face → |
| Q6_K ← GGUF · Abiray | 16.7 GB | 22.5 GB | fits · 1.5 GB spare | very close to the original | Hugging Face → |
| Q5_K_M GGUF · Abiray | 14.1 GB | 19.9 GB | fits · 4.1 GB spare | close; small differences in fine detail | Hugging Face → |
| Q4_K_M GGUF · Abiray | 11.6 GB | 17.4 GB | fits · 6.6 GB spare | good; some loss in fine detail and text | Hugging Face → |
| Q3_K_M GGUF · Abiray | 8.9 GB | 14.7 GB | fits · 9.3 GB spare | noticeable loss of detail | Hugging 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. Use the smallest encoder file; it still fits on its own.
04About this GPU
The RTX 4090 24 GB is an Ada Lovelace GPU with hardware FP8, so ComfyUI can compute Comfy-Org's FP8 files natively: small and fast. (Plain FP8 files use FP8 maths with the --fast fp8_matrix_mult option.)
RTX 4090 24 GB: 24 GB GDDR6X · 384-bit · 1008 GB/s · Ada Lovelace. Everything that runs on the RTX 4090 24 GB →
05What more VRAM would change
With 32 GB you could run MiniMax H3 Pruned at 8-bit or better (FP8, 21.0 GB) with no offloading: for example on the RTX 5090 32 GB.
What changes from the RTX 4090 24 GB to the RTX 5090 32 GB →
06Measured and reported results
| Label | Setup | Result | Peak VRAM | Date | Source |
|---|---|---|---|---|---|
| reported | minimax_h3_fl2va_pruned_int8_convrot + qwen3vl_32b nvfp4_awq · 832x480 · 8 steps “5 s clip (832×480, 8 steps): ~7 min” ~7 min approx = 420 s for 5 s clip; VRAM ~6.6 GB during sampling, ~22.5 GB spike at model load; 15 s clip ~25-30 min; no date shown | 420 s / clip | 22.5 GB | — | github.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 4090 24 GB?
Q6_K (16.7 GB) from Abiray/MiniMax-H3-Pruned-GGUF. Load it with Unet Loader (GGUF) from the ComfyUI-GGUF node pack.
Is FP8 faster than GGUF on the RTX 4090 24 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 4090 24 GB?
About 49.7 GB for the model file, text encoder and VAE listed on this page, and the same again free on disk. With a 24 GB card, 64 GB of system RAM or more is recommended.