MiniMax H3 Pruned VRAM requirements

01The files, and how much VRAM each needs

“Needed” = file + 5 GB working memory + 0.8 GB system reserve.

02By amount of VRAM

03Best GPU for MiniMax H3 Pruned

The cheapest cards (by launch price) that run it well, and every card sorted by memory: best GPU for MiniMax H3 Pruned → Planning bigger images or longer clips? Open the calculator →

04By graphics card

On RTX 40/50 GPUs the FP8 file is preferred over Q8_0 when both fit (hardware FP8). All verdicts are calculated; see how the numbers work.

05Text encoder, VAE and other files

The files the official ComfyUI workflows load next to the model. Sizes read from Hugging Face (2026-09-25). Every GPU page for this model lists the exact set to download for that card, with the total.

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 16 GB the encoder itself is too big for VRAM, so let it run from system RAM: slower prompt encoding, same images.

06Where the files go in ComfyUI

Standard ComfyUI folders. After copying files, press R in ComfyUI (or restart it) to refresh the lists. Some uploads need their uploader's own loader node — see the notes above.

07AMD, Intel and NVIDIA: which file types are fast

Every file type loads on every listed GPU, so the memory verdicts apply to all of them. What differs is speed: FP8 maths needs RTX 40/50 or RX 9000 (with ROCm 6.4+ and PyTorch 2.7+); ComfyUI's INT8 maths runs on NVIDIA and AMD, not on Intel; GGUF is unpacked on the fly on any GPU, which costs some speed. NVFP4 files are fast only on RTX 50. AMD runs ComfyUI on Windows through ROCm, Intel through PyTorch XPU; some custom nodes are NVIDIA-only. Source: ComfyUI model_management.py. AMD and Intel guide →

08Measured and reported results

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.

09Training a LoRA for MiniMax H3 Pruned

No trainer documentation with a VRAM figure for this model was found yet. What the trainers say for other models →

10Every file tracked for MiniMax H3 Pruned

33B (HF API 33.1B) omni transformer generating video WITH native stereo audio, 4-15 s, 768p native (2K via upscale pass). Two denoisers: fl2va (text/first/last-frame-to-video+audio) and ref2va (up to 9 images/3 videos/3 audio refs); load one. 'Pruned' variants are smaller official Comfy-Org cuts (40.2 GB bf16 vs 66.3 GB). VAEs: minimax_h3_video_vae_fp16 5207808496 + minimax_h3_audio_vae_fp32 605254808. Native ComfyUI support (day-0, Aug 2026 blog); Comfy blog says smallest variants total 42.5 GB (from 123.6 GB). unsloth/MiniMax-H3-GGUF (~1.18M downloads, pruned only, quants named Q3_K/Q4_K/UD-*) is documented for stable-diffusion.cpp, not ComfyUI - not listed.