MiniMax H3 on Mac: Files, Software and Timings (2026)

MiniMax H3 runs on a Mac, but only on some Macs, in some countries, and slowly. The smallest Mac with a published 5-second clip is a 24 GB M4 Pro, which needed 47 minutes 58.7 seconds for a 1344×768 clip by streaming weights from the SSD (Argus README, 2026). A 48 GB M5 Pro rendered a 352×608 clip of the same length in 6 minutes 54 seconds (Bambushu README, 2026). The settings differ, so those two times are not a ranking. Read the license section first: it grants rights only outside the United States, the United Kingdom, the European Union and the Republic of Korea.

Can you use MiniMax H3 where you live?

Not everywhere. The MiniMax H3 Community License Agreement, dated August 2, 2026, defines the excluded territories in one sentence (LICENSE, 2026):

“Excluded Territories” means the European Union, the United Kingdom, the Republic of Korea and the United States of America.

The “Applicable Territory” means “worldwide, excluding the Excluded Territories”. The grant opens with “Solely within the Applicable Territory, we grant you a non-exclusive, non-transferable, royalty-free, limited license”, and section V.4 states: “You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.” The license text also says that anyone in an excluded territory who wants to deploy the models is welcome to contact MiniMax about a license. MiniMax's license Q&A calls the limit “not yet”, not “not ever”, and links an application form for organizations.

Two more terms matter before you build on it. A company whose commercial products generate more than 20 million US dollars in yearly revenue needs a separate written authorization from MiniMax. ComfyUI's H3 documentation states a broader rule: commercial use of locally generated outputs requires a MiniMax commercial license sold through Comfy. Settle that with MiniMax before commercial work.

The community quantizations repeat the same terms. The unsloth GGUF card says the license “defines an Applicable Territory and excludes some jurisdictions from it”, and the MLX Serve card tells readers to check that their jurisdiction permits using the files before downloading them. ModelFit is not a law firm: if you live or work in an excluded territory, ask MiniMax or a lawyer before you download anything.

What is MiniMax H3, and what do the open weights include?

MiniMax H3 is a video model that generates video and stereo audio together, at 24 FPS and up to 15 seconds (model card, 2026). It is not MiniMax M3, the company's separate language model, which Hugging Face tags as image-text-to-text (M3 card).

The open release is H3-Base. The card says the shorter side defaults to 768 pixels, and the 2K output of the hosted workflow comes from an H3-Regenerate-2K module that is “not yet open-sourced”. So the local weights stop at 768p. Inside are a 33B-parameter diffusion transformer, a Qwen3-VL-32B text encoder, a video VAE and an audio VAE. Two checkpoints exist: FL2VA for text-to-video plus first and last frames, and Ref2VA for reference images, videos and audio.

MiniMax's own server example uses four GPUs (--num-gpus 4). Neither the model card nor the two ComfyUI H3 documentation pages (overview and native workflows) mention Apple Silicon, so every Mac number in this guide comes from community authors.

Which Macs can run MiniMax H3?

Published runs show H3 finishing 5-second clips on 24 GB, 48 GB, 64 GB and 128 GB Macs, and a 1-second clip on a 16 GB MacBook Air. The sources behind this guide include no 32 GB Mac. ModelFit ran none of these tests: each row cites its author, and every row uses different settings, so read the tables for order of magnitude, never as a leaderboard. Dates are the day each page, README revision or comment was published, because most authors do not state the run date.

The MiniMax H3 model page lists every complete build with its RAM verdict, and the H3 GGUF page lists the exact GGUF files. For other open video models on the same hardware, see ModelFit's video model list and the local video generation guide.

16 GB and 24 GB Macs

On the 16 GB Air, clips of 56 frames and more ran out of GPU memory, and the README adds that macOS gives the GPU about 74% of RAM. The 24 GB runs stream weights instead of holding them: Argus keeps two DiT blocks resident at a time, each about 1.29 GB in BF16. It reports a 15.8 GB peak footprint for the BF16 run and about 12.0 GB maximum RSS for the INT8 run, but the two memory figures use different accounting, so do not compare them. NFE counts denoiser forward passes.

48 GB and 64 GB Macs

Bambushu's times include model loading, sampling, decoding and saving, each from one run in an existing session rather than a cold-start average. The rznotes h3.c table gives total generation time without a step count. The same article reports a peak of 18.88 GiB for a shorter 512×512, 22-frame, 20-step run, where the denoise phase alone took 32.89 seconds.

128 GB Macs

The Zenn times are the generation time the author reports, and the peak memory ran from 66.5 GB for the shortest clip to 76.6 GB at 226 frames. The 362-frame case failed under memory pressure and swapping, and its first diffusion step alone took 79 minutes, so the 15-second headline is not a promise for a 128 GB Mac. The ComfyUI commenter's native MLX path is a custom port of the sampler that no stock template ships, and the same comment lists about 16 minutes 45 seconds for a roughly 0.7 MP, 5-second clip as the recommended quality setting.

The h3.c README also lists two Ref2VA renders on a 128 GB M5 Max, finished in 74.58 and 76.99 seconds with about a 40.1 GB peak footprint (h3.c README, 2026). It gives no resolution, frame count or step count for them, so this guide does not use them as a speed figure.

If you are choosing a machine for video work, the best Mac for AI video generation guide covers the memory tiers across models.

What makes an H3 clip slow, or fail?

Frames, resolution, step count and memory strategy each move the time, and the sources show how far:

• Frames: on the M4 Max, going from 56 to 124 frames (about 2.2 times) took the time from 75.2 to 310.1 seconds, about four times. The same article reports 211.7 seconds for a demo clip at 608×352 and 124 frames, so single timings carry noise (Zenn, 2026).
• Resolution: at 56 frames, 608×352 took 75.2 seconds and 960×544 took 203.3 seconds (Zenn, 2026). Width and height must be multiples of 32, and frame counts snap upward to 5 + 17n (5, 22, 39, 56, 124) per the h3.c README.
• Turbo LoRA: a 4-step distillation adapter cut the 608×352, 56-frame clip from 142.2 seconds at 16 steps to 75.2 seconds at 4. The MLX Serve card measured 188 seconds of sampling at 16 steps against 95 seconds at 4 on an M4 Max, at 960×544 and 39 frames (MLX Serve card, 2026).
• Quantization: the MLX Serve card says its quantization buys footprint and no speed. Zenn's numbers agree: 4-bit took 75.2 seconds at a 67.6 GB peak and 8-bit took 80.0 seconds at 80.5 GB, same clip.
• SSD streaming: h3.c's --ssd-streaming cut tracked DiT storage from about 36.5 GiB to 2.0 GiB at 512 square on an M5 Max, and a warm 50-block forward took 1.35 versus 2.49 seconds (84% slower). The README adds that the 2.0 GiB figure covers only the DiT's tracked storage, and the OS, media buffers and output size still need headroom.

Which software runs MiniMax H3 on a Mac?

Four routes have Mac evidence, and a fifth is documented but unmeasured on a Mac.

Four caveats decide whether a route works on your machine:

• macOS 26: the MLX Serve builds that support H3 (26.8.2 and later) need Metal 4.0. The Zenn author could not start any of them on macOS 15, where every build failed with “Failed to load the default metallib”, and concluded that macOS 26 is a practical requirement. The rznotes runs were on macOS 26.5 and 26.6.1.
• ComfyUI 0.30.0 or later: ComfyUI's docs say the base H3 templates need ComfyUI 0.30.0+, and those pages never mention Apple Silicon. Both Mac write-ups added third-party nodes: rznotes installed the ComfyUI-AppleSilicon-FP8 node and patched its C++ standard and Metal language version for its own versions, and Bambushu launches with --reserve-vram 10 --cache-none --disable-smart-memory. For the general install, see ModelFit's ComfyUI on Mac guide.
• Text encoder: the template's default encoder is an NVFP4 file. Bambushu calls it CUDA-only and swaps in a GGUF encoder, while rznotes ran it on an M5 Pro through the custom node. Copy one write-up whole instead of mixing parts.
• Small projects: Argus's README says an AI agent completed the project with very little human involvement, and the MLX Serve brew tap is third-party, so read what you install.

What do the MiniMax H3 files weigh?

Sizes below are GiB from the Hugging Face file listings (bytes divided by 2^30) unless a line says GB. ComfyUI's docs list decimal GB. The official repository holds the original FL2VA and Ref2VA trees side by side, and the card tells you to scope the download, so take one tree:

That is about 134 GiB for one tree, matching the 134.1GiB total rznotes computed (file listing, 2026; rznotes h3.c, 2026). Nobody on a Mac needs all of it. ComfyUI's docs rank the diffusion-model files from pruned_w6a8 at 16 GB through pruned_int8_convrot and pruned_fp8_scaled at 21 GB each and pruned_bf16 at 40 GB to the full bf16 at 66 GB, and tell you to prefer pruned builds (ComfyUI docs, 2026). The rznotes ComfyUI set was four files and about 40GB: the pruned int8 diffusion model at 19.5GiB, an NVFP4 text encoder at 14.6GiB and the two VAEs, while the full Comfy-Org repository was about 465GB.

The MLX packs are one directory each, with the text encoder, VAEs and tokenizer included: MLX Serve's 4-bit FL2VA pack is about 38 GiB and the 8-bit pack about 65 GiB (Zenn, 2026; pack card, 2026).

Is there a MiniMax H3 GGUF, and does it run on a Mac?

Yes, with a catch: the files target a video runner. The unsloth GGUF repository holds eight FL2VA denoisers, six Ref2VA denoisers and two Qwen3-VL text encoders, and its card says it is “Not an official MiniMax product”. It documents sd-cli --mode vid_gen from stable-diffusion.cpp. H3 produces video and audio, so llama.cpp and Ollama text-model commands do not apply.

The card calls three flags non-optional: --mode vid_gen, --cfg-scale 1.0 because H3 is distilled and cfg-free, and --backend te=cpu, which keeps the 12 GB text encoder off the GPU. Its example runs 640×384, 25 frames and 4 steps and was written for “a single card”. It says nothing about Macs, and the sources behind this guide include no Mac timing for sd-cli.

Totals add the two VAEs, 4.85 GiB for video and 0.56 GiB for audio, which the card says to take from Comfy-Org/MiniMax-H3. The other rungs are UD-Q2_K_XL at 7.51 GiB, Q3_K at 8.16 GiB, UD-Q3_K_XL at 8.90 GiB, Q5_0 at 12.97 GiB and Q6_K at 15.45 GiB. The card pairs the Q2_K_M encoder with the two smallest denoisers and Q4_K_M with the rest (GGUF card, 2026).

The GGUF route with Mac timings is ComfyUI. Bambushu's workflow loads a Q5_K_M denoiser (14.1 GB in its README), a Qwen3-VL Q4_K_M encoder (14.6 GB), a Turbo LoRA and the two VAEs, about 37 GB of disk in total. Those files come from other repositories than unsloth's, so their sizes differ. Below 32 GB, the README switches to a Q3_K_M denoiser (8.8 GB) and replaces the 32B encoder with a 4B projection encoder on the CPU, and its 16 GB MacBook Air run sits in that section.

The wizard on the ModelFit homepage picks text models for your Mac; it does not rate video models, so treat the tables above as the reference for H3.

FAQ

Can a 16 GB Mac run MiniMax H3?

Barely. The only 16 GB run in the sources is an M1 MacBook Air that made a 22-frame clip (about 1 second) at 352×608 in 17 minutes with a 4-step LoRA, or 36 minutes with an 8-step one, in the README's light setup (a Q3_K_M GGUF and a 4B encoder). Clips of 56 frames and more ran out of GPU memory (Bambushu, 2026). The smallest Mac with a published 5-second clip has 24 GB.

Can I run MiniMax H3 with Ollama or llama.cpp?

No. Ollama and llama.cpp text-model commands do not apply, because H3 generates video and audio. The Mac routes in the sources are h3.c, Argus MLX streaming, MLX Serve and ComfyUI, and the GGUF files target stable-diffusion.cpp's sd-cli or a ComfyUI GGUF loader. For chat models on the same Mac, the ModelFit picker names the best fit.

How much disk space does MiniMax H3 need on a Mac?

The official FL2VA tree is about 134 GiB. Mac authors downloaded far less: about 40GB for the rznotes ComfyUI int8 set, about 37 GB for Bambushu's GGUF set, and roughly 24 to 42 GiB for the unsloth GGUF sets with VAEs. MLX Serve's 4-bit pack is about 38 GiB and its 8-bit pack about 65 GiB (Zenn, 2026).

Can a Mac make the full 15-second clip?

The model card says up to 15 seconds, but the one published local attempt at that length failed. At 608×352 with the 4-bit pack on a 128 GB M4 Max, 362 frames hit memory pressure and swapping, and the first diffusion step took 79 minutes. The same Mac finished a 9.4-second clip in 654.0 seconds (Zenn, 2026).

Is it legal to run MiniMax H3 in the US, UK or EU?

The license lists the United States, the United Kingdom, the European Union and the Republic of Korea as excluded territories and authorizes use only outside them. It invites anyone there to contact MiniMax about a license, and the Q&A links an application form for organizations. ModelFit is not a law firm, so ask MiniMax or a lawyer, and do not treat a quantized repo's card as different terms.

Sources

• MiniMax H3 model card, LICENSE and license Q&A, MiniMax, August 2026
• MiniMax M3 model card, MiniMax
• Official file listing, Hugging Face, read October 24, 2026
• ComfyUI MiniMax H3 documentation and Comfy-Org/MiniMax-H3 files
• h3.c README, antirez, pinned commit 8974cc0
• Argus BF16 README and Argus INT8 README, August 11, 2026
• Bambushu minimax-h3-mac README, commit 6bc2f75, October 22, 2026
• rznotes: h3.c on an M5 Pro 64GB and rznotes: ComfyUI on an M5 Pro 64GB, August 13, 2026
• Zenn: MiniMax H3 on an M4 Max 128GB, August 12, 2026
• ComfyUI issue 15315, Apple Silicon comment, August 10, 2026
• MLX Serve 4-bit FL2VA card and mlx-serve
• unsloth MiniMax-H3-GGUF and stable-diffusion.cpp