Files read on Hugging Face: 2026-10-24

MiniMax H3 GGUF files and Mac RAM

MiniMax H3 is an open-weights model that generates video with synchronized stereo audio from text, optional first and last frames, or reference media. The open release is H3-Base, and on a Mac it runs through community MLX, Metal and ComfyUI routes.

MiniMax · Updated 2026-10-24
Answer

MiniMax H3 runs on a Mac with 48 GB or more. A published run on M5 Pro · 48 GB used the ComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow) build (source, 2026-09-14).

Smallest Mac with a published run
48 GBM5 Pro · 48 GB · ComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)
Smallest complete download
35.45 GBUnsloth GGUF Q4_K denoiser with the Q4_K_M Qwen3-VL encoder
Files read
2026-10-24Hugging Face

MiniMax H3 Community License Agreement. Commercial use: restricted. Excluded territories: European Union, United Kingdom, Republic of Korea, United States of America. Separate prior written authorization is required when commercial products and services generate more than 20 million US dollars in yearly revenue.

MiniMax H3 is a video-and-audio generation model, not MiniMax M3, which is a separate text LLM.

Table 1

Which exact GGUF files make a complete build?

Only builds with every required component and an exact probed file size are shown.

Unsloth GGUF Q4_K denoiser with the Q4_K_M Qwen3-VL encoder

gguf · Q4_K · 35.45 GB complete

Mac routes: stable-diffusion.cpp

Exact component files for Unsloth GGUF Q4_K denoiser with the Q4_K_M Qwen3-VL encoder
FileComponentRequiredDownloadSource
minimax_h3_fl2va_pruned-Q4_K.ggufunsloth/MiniMax-H3-GGUFH3-Omni-TransformerditYes11.42 GBHugging FaceRead 2026-10-24
qwen3vl_32b_minimax_h3-Q4_K_M.ggufunsloth/MiniMax-H3-GGUFH3-Encoder (Qwen3-VL-32B)text_encoderYes18.22 GBHugging FaceRead 2026-10-24
vae/minimax_h3_video_vae_fp16.safetensorsComfy-Org/MiniMax-H3H3-VisualVAEvaeYes5.21 GBHugging FaceRead 2026-10-24
vae/minimax_h3_audio_vae_fp32.safetensorsComfy-Org/MiniMax-H3H3-AudioVAEaudioYes0.61 GBHugging FaceRead 2026-10-24

Build sources: Unsloth MiniMax-H3 GGUF card.

ComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)

gguf · Q5_K_M · 36.57 GB complete

Mac routes: ComfyUI

Exact component files for ComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)
FileComponentRequiredDownloadSource
MiniMax-H3-FL2VA-Pruned-Q5_K_M.ggufAbiray/MiniMax-H3-Pruned-GGUFH3-Omni-TransformerditYes14.07 GBHugging FaceRead 2026-10-24
qwen3vl-32B-MiniMax-H3-Q4_K_M.ggufrealrebelai/MiniMax-H3_GGUFsH3-Encoder (Qwen3-VL-32B)text_encoderYes14.58 GBHugging FaceRead 2026-10-24
vae/minimax_h3_video_vae_fp16.safetensorsComfy-Org/MiniMax-H3H3-VisualVAEvaeYes5.21 GBHugging FaceRead 2026-10-24
vae/minimax_h3_audio_vae_fp32.safetensorsComfy-Org/MiniMax-H3H3-AudioVAEaudioYes0.61 GBHugging FaceRead 2026-10-24
H3-PK-Parasyte-Turbo.safetensorsPlaguekind/H3-LoraTurbo LoRAotherNo2.12 GBHugging FaceRead 2026-10-24

Build sources: Bambushu minimax-h3-mac README.

Table 2

Does it fit your Mac's RAM?

Published names the exact measured machine. est. is a memory calculation. unknown means the run verdict is not established, not that the model fails.

Run verdicts by Mac memory tier
Mac RAMComplete buildDownloadRun verdictBasis
8 GBUnsloth GGUF Q4_K denoiser with the Q4_K_M Qwen3-VL encoderQ4_K · stable-diffusion.cpp35.45 GBDoes not fitest.
16 GBUnsloth GGUF Q4_K denoiser with the Q4_K_M Qwen3-VL encoderQ4_K · stable-diffusion.cpp35.45 GBDoes not fitest.
24 GBUnsloth GGUF Q4_K denoiser with the Q4_K_M Qwen3-VL encoderQ4_K · stable-diffusion.cpp35.45 GBDoes not fitest.
32 GBUnsloth GGUF Q4_K denoiser with the Q4_K_M Qwen3-VL encoderQ4_K · stable-diffusion.cpp35.45 GBDoes not fitest.
48 GBComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)Q5_K_M · ComfyUI36.57 GBRuns · PublishedPublishedApple M5 Pro · 48 GB RAM
64 GBComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)Q5_K_M · ComfyUI36.57 GBRuns · Published on M5 Pro · 48 GBPublishedApple M5 Pro · 48 GB RAM
96 GBComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)Q5_K_M · ComfyUI36.57 GBRuns · Published on M5 Pro · 48 GBPublishedApple M5 Pro · 48 GB RAM
128 GBComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)Q5_K_M · ComfyUI36.57 GBRuns · Published on M5 Pro · 48 GBPublishedApple M5 Pro · 48 GB RAM

How does it rank in blind votes?

#8 of 48 in text to video blind votes, from 11,417 votes.

Quality: blind human votes from minimax-h3 (arena.ai), CC BY 4.0, snapshot 2026-09-22. Scores are for the Arena checkpoint, not your local quantized build.

Which Mac software routes are listed?

  • ComfyUI

    Official

    Formats: safetensors, gguf. Memory policy: component stages stay resident.

    On Apple Silicon ComfyUI runs on PyTorch MPS, not MLX. GGUF files need the separate city96/ComfyUI-GGUF custom node. Resident size is the sum of the weights, an upper bound: ComfyUI can unload models between stages and publishes no per-model Mac peak.

    Sources: ComfyUI README · ComfyUI system requirements · ComfyUI CLI arguments · ComfyUI-GGUF README

  • stable-diffusion.cpp

    Official

    Formats: gguf, safetensors. Memory policy: component stages stay resident.

    C/C++ runner with Metal on Mac and CUDA on NVIDIA. It reads GGUF and safetensors, and its docs keep a page per model family. Resident size is the sum of the weights, an upper bound.

    Sources: stable-diffusion.cpp README, backends

What does the official configuration say?

"The official SGLang example runs on 4 GPUs (--num-gpus 4). No official MiniMax Mac memory figure was found."

Source: MiniMax-H3 model card · Read 2026-10-24.

What known issues should you check?

  • Known issue 1

    The open release is H3-Base. H3-Context-IR and H3-Regenerate-2K are not included, so the local checkpoint is not the complete hosted 2K workflow.

    Source: MiniMax-H3 model card · Read 2026-10-24

  • Known issue 2

    A 362-frame 608x352 run with the 4-bit pack on a 128 GB M4 Max failed under memory pressure and swapping; the first diffusion step alone took 79 minutes.

    Source: Zenn: MiniMax H3 on M4 Max 128GB · Read 2026-10-24

  • Known issue 3

    Argus reports peak memory footprint for the BF16 run and maximum RSS for the INT8 run, and says Metal memory follows different accounting than RSS, so the two figures are not directly comparable.

    Source: Argus MLX route README (BF16 run) · Read 2026-10-24

Sources