MiniMax H3 on Mac: files, RAM and published times
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 H3 runs on a Mac with 24 GB or more. A published run on M4 Pro · 24 GB used the Official FL2VA BF16 (MiniMaxAI/MiniMax-H3) build (source, 2026-08-11).
- Smallest Mac with a published run
- 24 GBM4 Pro · 24 GB · Official FL2VA BF16 (MiniMaxAI/MiniMax-H3)
- 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.
See the exact GGUF files and sizes
MiniMax H3 is a video-and-audio generation model, not MiniMax M3, which is a separate text LLM.
Get the weekly local AI newsletter
The Thursday ModelFit newsletter: new local models and what fits your machine.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
What do you download?
Only builds with every required component and an exact probed file size are shown.
MLX Serve FL2VA 4-bit pack with Turbo LoRA
mlx · 4-bit · 41.10 GB complete · RecommendedMac routes: ddalcu/mlx-serve
| File | Component | Required | Download | Source |
|---|---|---|---|---|
transformer.safetensorsddalcu/MiniMax-H3-FL2VA-MLX-Serve-4bit | H3-Omni-Transformerdit | Yes | 18.70 GB | Hugging FaceRead 2026-10-24 |
text_encoder.safetensorsddalcu/MiniMax-H3-FL2VA-MLX-Serve-4bit | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 15.80 GB | Hugging FaceRead 2026-10-24 |
video_vae.safetensorsddalcu/MiniMax-H3-FL2VA-MLX-Serve-4bit | H3-VisualVAEvae | Yes | 5.21 GB | Hugging FaceRead 2026-10-24 |
audio_vae.safetensorsddalcu/MiniMax-H3-FL2VA-MLX-Serve-4bit | H3-AudioVAEaudio | Yes | 0.61 GB | Hugging FaceRead 2026-10-24 |
turbo_lora.safetensorsddalcu/MiniMax-H3-FL2VA-MLX-Serve-4bit | Turbo LoRAother | No | 0.78 GB | Hugging FaceRead 2026-10-24 |
Build sources: MLX Serve 4-bit pack card.
Unsloth GGUF Q4_K denoiser with the Q4_K_M Qwen3-VL encoder
gguf · Q4_K · 35.45 GB completeMac routes: stable-diffusion.cpp
| File | Component | Required | Download | Source |
|---|---|---|---|---|
minimax_h3_fl2va_pruned-Q4_K.ggufunsloth/MiniMax-H3-GGUF | H3-Omni-Transformerdit | Yes | 11.42 GB | Hugging FaceRead 2026-10-24 |
qwen3vl_32b_minimax_h3-Q4_K_M.ggufunsloth/MiniMax-H3-GGUF | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 18.22 GB | Hugging FaceRead 2026-10-24 |
vae/minimax_h3_video_vae_fp16.safetensorsComfy-Org/MiniMax-H3 | H3-VisualVAEvae | Yes | 5.21 GB | Hugging FaceRead 2026-10-24 |
vae/minimax_h3_audio_vae_fp32.safetensorsComfy-Org/MiniMax-H3 | H3-AudioVAEaudio | Yes | 0.61 GB | Hugging 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 completeMac routes: ComfyUI
| File | Component | Required | Download | Source |
|---|---|---|---|---|
MiniMax-H3-FL2VA-Pruned-Q5_K_M.ggufAbiray/MiniMax-H3-Pruned-GGUF | H3-Omni-Transformerdit | Yes | 14.07 GB | Hugging FaceRead 2026-10-24 |
qwen3vl-32B-MiniMax-H3-Q4_K_M.ggufrealrebelai/MiniMax-H3_GGUFs | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 14.58 GB | Hugging FaceRead 2026-10-24 |
vae/minimax_h3_video_vae_fp16.safetensorsComfy-Org/MiniMax-H3 | H3-VisualVAEvae | Yes | 5.21 GB | Hugging FaceRead 2026-10-24 |
vae/minimax_h3_audio_vae_fp32.safetensorsComfy-Org/MiniMax-H3 | H3-AudioVAEaudio | Yes | 0.61 GB | Hugging FaceRead 2026-10-24 |
H3-PK-Parasyte-Turbo.safetensorsPlaguekind/H3-Lora | Turbo LoRAother | No | 2.12 GB | Hugging FaceRead 2026-10-24 |
Build sources: Bambushu minimax-h3-mac README.
ComfyUI pruned INT8 convrot denoiser with the NVFP4 encoder (rznotes setup)
safetensors · INT8 convrot · 42.47 GB completeMac routes: ComfyUI
| File | Component | Required | Download | Source |
|---|---|---|---|---|
diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensorsComfy-Org/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 20.97 GB | Hugging FaceRead 2026-10-24 |
text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensorsComfy-Org/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 15.69 GB | Hugging FaceRead 2026-10-24 |
vae/minimax_h3_video_vae_fp16.safetensorsComfy-Org/MiniMax-H3 | H3-VisualVAEvae | Yes | 5.21 GB | Hugging FaceRead 2026-10-24 |
vae/minimax_h3_audio_vae_fp32.safetensorsComfy-Org/MiniMax-H3 | H3-AudioVAEaudio | Yes | 0.61 GB | Hugging FaceRead 2026-10-24 |
Build sources: rznotes: MiniMax H3 with ComfyUI on an M5 Pro 64GB.
Argus calibrated INT8 DiT with the official text encoder and VAEs
mlx · INT8 · 115.53 GB completeMac routes: Argus-AiTeam/minimax-h3-mac (MLX streaming)
| File | Component | Required | Download | Source |
|---|---|---|---|---|
model-00001-of-00008.safetensorswater1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 | H3-Omni-Transformerdit | Yes | 5.35 GB | Hugging FaceRead 2026-10-24 |
model-00002-of-00008.safetensorswater1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 | H3-Omni-Transformerdit | Yes | 5.35 GB | Hugging FaceRead 2026-10-24 |
model-00003-of-00008.safetensorswater1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 | H3-Omni-Transformerdit | Yes | 5.27 GB | Hugging FaceRead 2026-10-24 |
model-00004-of-00008.safetensorswater1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 | H3-Omni-Transformerdit | Yes | 5.36 GB | Hugging FaceRead 2026-10-24 |
model-00005-of-00008.safetensorswater1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 | H3-Omni-Transformerdit | Yes | 5.35 GB | Hugging FaceRead 2026-10-24 |
model-00006-of-00008.safetensorswater1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 | H3-Omni-Transformerdit | Yes | 5.27 GB | Hugging FaceRead 2026-10-24 |
model-00007-of-00008.safetensorswater1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 | H3-Omni-Transformerdit | Yes | 5.34 GB | Hugging FaceRead 2026-10-24 |
model-00008-of-00008.safetensorswater1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 | H3-Omni-Transformerdit | Yes | 0.51 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00001-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.93 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00002-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00003-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00004-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00005-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00006-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00007-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00008-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00009-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00010-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00011-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00012-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00013-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00014-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 3.27 GB | Hugging FaceRead 2026-10-24 |
FL2VA/video_vae/source/model.safetensorsMiniMaxAI/MiniMax-H3 | H3-VisualVAEvae | Yes | 10.42 GB | Hugging FaceRead 2026-10-24 |
FL2VA/audio_vae/model.safetensorsMiniMaxAI/MiniMax-H3 | H3-AudioVAEaudio | Yes | 0.61 GB | Hugging FaceRead 2026-10-24 |
Build sources: Argus calibrated INT8 DiT card · Argus MLX route README.
Official FL2VA BF16 (MiniMaxAI/MiniMax-H3)
safetensors · BF16 · 144.02 GB completeMac routes: Argus-AiTeam/minimax-h3-mac (MLX streaming), antirez/h3.c (Metal)
| File | Component | Required | Download | Source |
|---|---|---|---|---|
FL2VA/transformer/model-00001-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.23 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00002-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00003-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00004-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00005-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00006-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00007-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00008-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00009-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00010-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00011-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00012-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 5.16 GB | Hugging FaceRead 2026-10-24 |
FL2VA/transformer/model-00013-of-00013.safetensorsMiniMaxAI/MiniMax-H3 | H3-Omni-Transformerdit | Yes | 4.24 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00001-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.93 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00002-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00003-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00004-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00005-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00006-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00007-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00008-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00009-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00010-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00011-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00012-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00013-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 4.88 GB | Hugging FaceRead 2026-10-24 |
FL2VA/text_encoder/model-00014-of-00014.safetensorsMiniMaxAI/MiniMax-H3 | H3-Encoder (Qwen3-VL-32B)text_encoder | Yes | 3.27 GB | Hugging FaceRead 2026-10-24 |
FL2VA/video_vae/source/model.safetensorsMiniMaxAI/MiniMax-H3 | H3-VisualVAEvae | Yes | 10.42 GB | Hugging FaceRead 2026-10-24 |
FL2VA/audio_vae/model.safetensorsMiniMaxAI/MiniMax-H3 | H3-AudioVAEaudio | Yes | 0.61 GB | Hugging FaceRead 2026-10-24 |
Build sources: MiniMax-H3 model card · Argus MLX route README.
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.
| Mac RAM | Complete build | Download | Run verdict | Basis |
|---|---|---|---|---|
| 8 GB | MLX Serve FL2VA 4-bit pack with Turbo LoRA4-bit · ddalcu/mlx-serve | 41.10 GB | Does not fit | est. |
| 16 GB | MLX Serve FL2VA 4-bit pack with Turbo LoRA4-bit · ddalcu/mlx-serve | 41.10 GB | Does not fit | est. |
| 24 GB | Argus calibrated INT8 DiT with the official text encoder and VAEsINT8 · Argus-AiTeam/minimax-h3-mac (MLX streaming) | 115.53 GB | Runs · Published | PublishedMacBook Pro (Mac16,8) · Apple M4 Pro · 24 GB RAM |
| 32 GB | Argus calibrated INT8 DiT with the official text encoder and VAEsINT8 · Argus-AiTeam/minimax-h3-mac (MLX streaming) | 115.53 GB | Runs · Published on M4 Pro · 24 GB | PublishedMacBook Pro (Mac16,8) · Apple M4 Pro · 24 GB RAM |
| 48 GB | ComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)Q5_K_M · ComfyUI | 36.57 GB | Runs · Published | PublishedApple M5 Pro · 48 GB RAM |
| 64 GB | ComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)Q5_K_M · ComfyUI | 36.57 GB | Runs · Published on M5 Pro · 48 GB | PublishedApple M5 Pro · 48 GB RAM |
| 96 GB | ComfyUI GGUF Q5_K_M pruned denoiser with Parasyte Turbo (Bambushu workflow)Q5_K_M · ComfyUI | 36.57 GB | Runs · Published on M5 Pro · 48 GB | PublishedApple M5 Pro · 48 GB RAM |
| 128 GB | MLX Serve FL2VA 4-bit pack with Turbo LoRA4-bit · ddalcu/mlx-serve | 41.10 GB | Runs · Published | PublishedMacBook Pro · Apple M4 Max · 128 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.
What published Mac times are available?
Third-party results keep their original machine and settings. ModelFit does not run these benchmarks.
| Machine and RAM | Settings | App | Result | Source |
|---|---|---|---|---|
| MacBook Air · Apple M1 · 16 GB RAM | 352x608 · 4 steps · 22 frames · H3 Turbo LoRA merged in, 4 steps, low_vram on. Light setup: Q3_K_M GGUF denoiser, ClipProj 4B encoder on the CPU, audio enabled. The date is the README commit date; the run date is not stated. | ComfyUI | Success17 minPeak memory unknown | Bambushu minimax-h3-mac READMEPublished 2026-10-22 · Grade B |
| MacBook Air · Apple M1 · 16 GB RAM | 352x608 · 8 steps · 22 frames · Parasyte LoRA, 8 steps. Light setup: Q3_K_M GGUF denoiser, ClipProj 4B encoder on the CPU, audio enabled. The date is the README commit date; the run date is not stated. | ComfyUI | Success36 minPeak memory unknown | Bambushu minimax-h3-mac READMEPublished 2026-10-22 · Grade B |
| MacBook Air · Apple M1 · 16 GB RAM | 352x608 · 56 frames · 56 frames and more. Light setup: Q3_K_M GGUF denoiser, ClipProj 4B encoder on the CPU, audio enabled. The date is the README commit date; the run date is not stated. | ComfyUI | FailedTime unknownPeak memory unknown | Bambushu minimax-h3-mac READMEPublished 2026-10-22 · Grade B |
| Apple M5 Pro · 48 GB RAM | 352x608 · 5.2 s · 8 steps · 124 frames · 24 fps · Q5_K_M · Pruned FL2VA GGUF Q5_K_M with Qwen3-VL 32B Q4_K_M, Parasyte Turbo strength 1.5, ER-SDE beta57, sigma shift video 12 and audio 3, audio enabled, seed 77001. Time includes model loading, sampling, decoding and saving; one run in an existing session. PyTorch 2.13.0, ComfyUI v0.34.0 plus 11 commits. | ComfyUI | Success6m 54sPeak memory unknown | Bambushu minimax-h3-mac READMEPublished 2026-09-14 · Grade A |
| Apple M5 Pro · 48 GB RAM | 480x864 · 5.2 s · 8 steps · 124 frames · 24 fps · Q5_K_M · Pruned FL2VA GGUF Q5_K_M with Qwen3-VL 32B Q4_K_M, Parasyte Turbo strength 1.5, ER-SDE beta57, sigma shift video 12 and audio 3, audio enabled, seed 77001. Time includes model loading, sampling, decoding and saving; one run in an existing session. PyTorch 2.13.0, ComfyUI v0.34.0 plus 11 commits. | ComfyUI | Success12m 52sPeak memory unknown | Bambushu minimax-h3-mac READMEPublished 2026-09-14 · Grade A |
| MacBook Pro · Apple M5 Pro · 64 GB RAM | 352x608 · 5.2 s · 10 steps · 124 frames · INT8 convrot · First and last frame image-to-video with the model audio track. ComfyUI v0.31.0 with the ComfyUI-AppleSilicon-FP8 custom node, macOS 26.5. The date is the article date; the run date is not stated. | ComfyUI | Success501.6 sPeak memory unknown | rznotes: MiniMax H3 with ComfyUI on an M5 Pro 64GBPublished 2026-08-13 · Grade A |
| MacBook Pro · Apple M5 Pro · 64 GB RAM | 576x1024 · 5.2 s · 124 frames · BF16 · Total generation time with the full BF16 checkpoint; the step count is not stated. h3.c commit 03cb133, macOS 26.6.1. The date is the article date; the run date is not stated. | antirez/h3.c (Metal) | Success1,859.1 sPeak memory unknown | rznotes: h3.c on an M5 Pro 64GBPublished 2026-08-13 · Grade A |
| MacBook Pro · Apple M5 Pro · 64 GB RAM | 576x1024 · 5.2 s · 124 frames · int8 at load (h3.c) · Same checkpoint as the BF16 run with the h3.c int8 path, so it is not a separate download and is not tied to a build. Total generation time; the step count is not stated. h3.c commit 03cb133, macOS 26.6.1. The date is the article date. | antirez/h3.c (Metal) | Success1,440 sPeak memory unknown | rznotes: h3.c on an M5 Pro 64GBPublished 2026-08-13 · Grade B |
| MacBook Pro · Apple M4 Max · 128 GB RAM | 384x224 · 2.3 s · 4 steps · 56 frames · 4-bit · Turbo LoRA, audio enabled, same prompt and seed. Generation time as reported; the run date is not stated. | ddalcu/mlx-serve | Success31.2 s66.7 GB peak | Zenn: MiniMax H3 on M4 Max 128GBPublished 2026-08-12 · Grade A |
| MacBook Pro · Apple M4 Max · 128 GB RAM | 512x288 · 2.3 s · 4 steps · 56 frames · 4-bit · Turbo LoRA, audio enabled, same prompt and seed. Generation time as reported; the run date is not stated. | ddalcu/mlx-serve | Success53.8 s68.6 GB peak | Zenn: MiniMax H3 on M4 Max 128GBPublished 2026-08-12 · Grade A |
| MacBook Pro · Apple M4 Max · 128 GB RAM | 608x352 · 5.2 s · 4 steps · 124 frames · 4-bit · Turbo LoRA, audio enabled, same prompt and seed. Generation time as reported; the run date is not stated. | ddalcu/mlx-serve | Success310.1 s70.4 GB peak | Zenn: MiniMax H3 on M4 Max 128GBPublished 2026-08-12 · Grade A |
| MacBook Pro · Apple M4 Max · 128 GB RAM | 608x352 · 9.4 s · 4 steps · 226 frames · 4-bit · Turbo LoRA, audio enabled, same prompt and seed. Generation time as reported; the run date is not stated. | ddalcu/mlx-serve | Success654 s76.6 GB peak | Zenn: MiniMax H3 on M4 Max 128GBPublished 2026-08-12 · Grade A |
| MacBook Pro · Apple M4 Max · 128 GB RAM | 608x352 · 0.9 s · 4 steps · 22 frames · 4-bit · Turbo LoRA, audio enabled, same prompt and seed. Generation time as reported; the run date is not stated. | ddalcu/mlx-serve | Success33.3 s66.5 GB peak | Zenn: MiniMax H3 on M4 Max 128GBPublished 2026-08-12 · Grade A |
| MacBook Pro · Apple M4 Max · 128 GB RAM | 608x352 · 4 steps · 362 frames · 4-bit · Turbo LoRA. The run swapped under memory pressure and the first diffusion step alone took 79 minutes. | ddalcu/mlx-serve | FailedTime unknownPeak memory unknown | Zenn: MiniMax H3 on M4 Max 128GBPublished 2026-08-12 · Grade B |
| MacBook Pro · Apple M4 Max · 128 GB RAM | 608x352 · 2.3 s · 4 steps · 56 frames · 4-bit · Turbo LoRA, audio enabled, same prompt and seed. Generation time as reported; the run date is not stated. | ddalcu/mlx-serve | Success75.2 s67.6 GB peak | Zenn: MiniMax H3 on M4 Max 128GBPublished 2026-08-12 · Grade A |
| MacBook Pro · Apple M4 Max · 128 GB RAM | 768x448 · 2.3 s · 4 steps · 56 frames · 4-bit · Turbo LoRA, audio enabled, same prompt and seed. Generation time as reported; the run date is not stated. | ddalcu/mlx-serve | Success114.3 s72.8 GB peak | Zenn: MiniMax H3 on M4 Max 128GBPublished 2026-08-12 · Grade A |
| MacBook Pro · Apple M4 Max · 128 GB RAM | 960x544 · 2.3 s · 4 steps · 56 frames · 4-bit · Turbo LoRA, audio enabled, same prompt and seed. Generation time as reported; the run date is not stated. | ddalcu/mlx-serve | Success203.3 s75.6 GB peak | Zenn: MiniMax H3 on M4 Max 128GBPublished 2026-08-12 · Grade A |
| MacBook Pro (Mac16,8) · Apple M4 Pro · 24 GB RAM | 1344x768 · 5.2 s · 5 steps · 124 frames · 24 fps · BF16 · Five sigma points are four DiT forwards. LightX2V 4-step 768p Turbo adapter, two streamed blocks, no block cache, audio enabled. Peak memory footprint is approximate. The date is the source revision date; the run date is not stated. | Argus-AiTeam/minimax-h3-mac (MLX streaming) | Success2,878.7 s15.8 GB peak | Argus MLX route README (BF16 run)Published 2026-08-11 · Grade A |
| MacBook Pro (Mac16,8) · Apple M4 Pro · 24 GB RAM | 768x448 · 5.2 s · 9 steps · 124 frames · 24 fps · INT8 · Nine sigma points are eight DiT forwards. Calibrated INT8 DiT, text encoder streamed in 4-bit, native BF16 MLX Turbo, streaming and no block cache. Peak is maximum RSS, approximate. The date is the source revision date; the run date is not stated. | Argus-AiTeam/minimax-h3-mac (MLX streaming) | Success1,378.0 s12.0 GB | Argus MLX route README (INT8 run)Published 2026-08-11 · Grade A |
Which Mac software routes are listed?
Argus-AiTeam/minimax-h3-mac (MLX streaming)
CommunityFormats: safetensors, mlx. Memory policy: DiT weights can stream from SSD.
Sources: Argus MLX route README
ComfyUI
OfficialFormats: 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
antirez/h3.c (Metal)
CommunityFormats: safetensors. Memory policy: DiT weights can stream from SSD.
The ssd-streaming policy applies only with the --ssd-streaming flag; the default keeps the DiT resident.
Sources: h3.c README
ddalcu/mlx-serve
CommunityFormats: mlx. Memory policy: component stages stay resident.
Sources: MLX Serve 4-bit pack card
stable-diffusion.cpp
OfficialFormats: 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.
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
Frequently asked questions
Can MiniMax H3 run on a 24 GB Mac?
Yes, with block streaming. Argus MLX ran the BF16 weights on a 24 GB M4 Pro at 1344x768 and 124 frames in 2,878.7 s, and the calibrated INT8 DiT ran 768x448 in 1,378.0 s. Both are end-to-end times for a clip of about 5 seconds.
What is the smallest Mac that ran H3?
A 16 GB M1 MacBook Air made a 22-frame 352x608 clip in 17 minutes with a 4-step Turbo LoRA, or 36 minutes with an 8-step LoRA, in Bambushu's light ComfyUI setup. Clips of 56 frames and more ran out of GPU memory.
Is MiniMax H3 licensed for commercial use?
Not everywhere. The MiniMax H3 Community License excludes the European Union, the United Kingdom, the Republic of Korea and the United States of America, and asks for separate written authorization above 20 million US dollars in yearly revenue.
Sources
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24
Read 2026-10-24