MiniMax H3 in ComfyUI: Files, VRAM and the Licence That Excludes the US

MiniMax H3 runs in ComfyUI on core nodes from four files totalling 40.07 GB, or 42.03 GB with the Turbo LoRA the stock templates also ask for. Neither MiniMax nor ComfyUI publishes a VRAM figure for it. The nearest official guidance, in the diffusers documentation, is a 24 to 32 GB card streaming weights from system RAM, or 12 to 16 GB at a smaller canvas. And before you download anything: the licence grants no rights in the United States, the European Union, the United Kingdom or the Republic of Korea.

This guide lists the exact files each local template loads, the smallest download, the settings, the Turbo LoRAs, the errors people report and what the licence says. BFWR has not run MiniMax H3. Everything here comes from MiniMax’s model card and licence, ComfyUI’s documentation, templates and source code, the published file listings and GitHub issue reports. GB means a billion bytes throughout.

Quick answer

• Licence: no rights in the US, EU, UK or South Korea, including for using or displaying the generated videos there. Elsewhere, commercial use is allowed below US$20M yearly revenue.
• Stock T2V and I2V files: minimax_h3_fl2va_pruned_int8_convrot.safetensors, qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, minimax_h3_video_vae_int8_convrot.safetensors, minimax_h3_audio_vae_fp32.safetensors.
• Smallest set from Comfy-Org: 35.09 GB, with the pruned_w6a8 transformer on ComfyUI 0.38.0 or later.
• Settings: res_multistep, simple, 20 steps, no negative prompt, 1344×768 native. Turbo: 8 steps (4 for reference-to-video).
• Most common fix: the int8 files need PyTorch built for CUDA 13.0.

The licence: read this before you download

Hugging Face shows license:other. The text is the MiniMax H3 Community License Agreement, dated 2 August 2026. This is a reading of the published text, not legal advice.

• Four territories are excluded. The grant applies “Solely within the Applicable Territory”, which is worldwide minus the Excluded Territories: “the European Union, the United Kingdom, the Republic of Korea and the United States of America”. Tencent’s HunyuanVideo 1.5 licence excludes the EU, UK and South Korea; H3 adds the US.
• The outputs are covered too. Section V.4: “You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.” As written, a clip made in a licensed country and displayed in Germany falls outside the licence. How that applies to a video posted worldwide is a question for a lawyer.
• There is a route for those territories. Section II: “should any person in such Excluded Territories be interested in deploying our models, you are welcome to contact us about obtaining a license, which will be granted based on robust controls and guardrails”. MiniMax’s request form asks for a legal company name and business email. Comfy also sells commercial licences, with its Professional tier listed from $5,000 a month on 9 November 2026.
• Commercial use elsewhere. There is no non-commercial clause. Section IV requires written authorisation above US$20 million in yearly revenue from your commercial products and services, and “MiniMax H3” shown prominently on the interface of a commercial product that uses it.
• ComfyUI’s two pages disagree. The tutorial says “Commercial use of locally generated outputs requires a MiniMax commercial license”. Comfy’s licence page says “The H3 Community License generally permits commercial use, provided your commercial products and services generate no more than $20 million in yearly revenue and you comply with its other terms.” The second matches the licence text.
• Outputs are yours, with limits. “MiniMax claims no rights over the Outputs you generate.” But Section V.3 bars using them to improve any AI model other than H3 and its derivatives, and policy item 12 forbids publishing generated content “without clearly and prominently disclosing that such information and/or content is machine-generated”. Item 19 bans military use.

If you offer H3 to others, you must also pass on the licence, bind users to its restrictions and run safeguards. Neither Hugging Face repository is gated, so a download will succeed from anywhere; that says nothing about whether you are licensed. Our Hunyuan3D 2.1 guide covers a similar territory clause.

What MiniMax H3 is

H3 generates video with native stereo audio: 4 to 15 seconds, 24 FPS, 32 kHz, 768 pixels on the short side. The Hugging Face API counts 33.12B parameters. There are two checkpoints: FL2VA for text-to-video and first/last-frame image-to-video, and Ref2VA for up to 9 reference images, 3 videos and 3 audio clips.

H3-Context-IR, the prompt processor MiniMax strongly recommends, is hosted only, and H3-Regenerate-2K had no files in MiniMax’s repository on 9 November. ComfyUI’s tutorial is plain about it: “2K is a hosted-service output and needs a separate upscale pass in ComfyUI.”

Which files each local template loads

These come from each template’s JSON. All live at Comfy-Org/MiniMax-H3 unless noted.

All of them share the encoder qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (loaded by CLIPLoader with type minimax) and the two VAEs. Despite the name, Comfy-Org’s card says “This nvfp4 text encoder does not require Blackwell GPU to use.”

ComfyUI/models/
├── diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
├── text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
├── vae/minimax_h3_video_vae_int8_convrot.safetensors
├── vae/minimax_h3_audio_vae_fp32.safetensors
└── loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors

The Turbo LoRA is off by default, but ComfyUI’s tutorial lists it as “Required by the workflow’s model scan”, so the template’s missing-model dialog will include it.

The real download size, and the smallest set

ComfyUI’s template catalogue lists 56.91 GB for T2V and 44.45 GB for I2V and R2V, but the files those templates load add up to 42.03 GB.

Use a pruned transformer. MiniMax’s card says about 13B of the parameters sit in AdaLN branches that inference does not need, and the full BF16 file is 26.05 GB larger, which is 13B at two bytes each. The minimax_h3_fl2va_pruned_w6a8.safetensors file needs ComfyUI 0.38.0 or later, and Kijai’s notes call the format a work in progress. Comfy-Org’s own guidance: “For diffusion models prefer int8_convrot if you are able to use pytorch with cu130.”

How much VRAM and RAM it needs

None of MiniMax’s card, ComfyUI’s six H3 pages or Comfy-Org’s README states a VRAM figure, so this is file arithmetic and other projects’ guidance, not a measurement. The stock files load in stages: the 15.69 GB encoder, the 20.97 GB transformer, then the VAEs. Their 40.07 GB sum is not a VRAM requirement, because they need not all sit on the GPU at once, and it is not a ceiling, because file sizes exclude the memory used while sampling.

Other official sources, for different software:

• diffusers: on 24 to 32 GB cards, quantize to int8 and stream the transformer from RAM. “On 12 to 16 GB the same recipe works with the video VAE group offloaded too” at a canvas such as 960×544. “Expect the weights to live in host RAM: around 75 GB of it at int8.” That is for the unpruned weights.
• SGLang: its RTX 4090 recipe, streaming the transformer and encoder, reports “GPU peak stays about 18 GB”.

ComfyUI streams weights with Dynamic VRAM. Flags from cli_args.py: --lowvram “Doesn’t do anything if dynamic vram is enabled”, --reserve-vram sets memory aside for other software, and --disable-dynamic-vram returns to estimate-based loading.

User reports vary and are not guarantees. A ComfyUI collaborator says one reference workflow “runs on my 4090 as it is just fine”; a user on #16150 ran a W4A8 model on 8 GB of VRAM and 16 GB of RAM; but a 32 GB RTX 5090 with 64 GB of RAM ran out of memory on #16705. For choosing a card, see our GPU comparison.

Settings in the stock templates

Keep width and height to multiples of 32; length snaps to the 17k+5 grid. Prompts follow a structure: integrated_multimodal_description: with timed [Shot 1] entries, then overall_soundscape: and non_diegetic_music:, with spoken lines in <d> tags. R2V inputs are <Picture 1>, <Video 1> and <Audio 1>. See ComfyUI’s prompt guide.

ComfyUI listens only on 127.0.0.1 by default. Leave --listen off unless you mean to expose it.

Turbo LoRAs, FastH3 and Pika H3 Flex

In the T2V and I2V templates, switch on turbo_mode; turbo_steps defaults to 8 and turbo_model_strength to 1.0. ComfyUI calls the trade “faster, slightly softer motion/audio”. For R2V it suggests 20 to 25 steps when references must hold.

ComfyUI’s default shift is 12 for video and 3 for audio, matching the 544p-trained LoRAs. To match ModelTC’s values for a 768p LoRA, add ModelSamplingMiniMaxH3 and set shift_video to 6 and shift_audio to 3.

FastH3 is a distilled checkpoint: exactly 8 steps, no reference-to-video, and ComfyUI warns it “trades some motion and audio fidelity for speed”. Pika H3 Flex is a 1.96 GB LoRA (pika_h3_flex_fl2va_v1_bf16.safetensors) needing ComfyUI 0.44.0, run with the lcm sampler at six steps.

Local templates against cloud templates

The template browser mixes the two. Of 28 H3-family templates on 9 November, 15 run locally and 13 begin with api_, including everything called H3 Max. The H3 API nodes carry a price badge in ComfyUI’s source, use none of the files above, and send your prompt off your machine. Our templates guide shows how to tell them apart.

Common failures

All of these issues were open on 9 November, and most fixes are users’ reports, so treat them as things to try.

• Pure noise or a mosaic. On #15419, a ComfyUI collaborator wrote “Update to cu130 minimum to use the int8 models.” Another user traced the same noise to corrupt, unfinalised downloads, so compare file sizes against the table. AMD RDNA3 cards have their own thread, #15314.
• to() received an invalid combination of arguments - got (NestedTensor). You are using VAE Decode (Tiled). Use plain VAEDecode; a contributor on #15274 notes “The minimax VAE has build in tiling already.”
• Missing model dialog, or too many values to unpack (expected 3). The template checks exact filenames, so having only minimax_h3_video_vae_fp16.safetensors still flags the int8 VAE as missing (#16917). Download it or select your file in vae_name. The unpack error there came from the audio VAE sitting in the video VAE slot.
• UnicodeDecodeError or Unknown quantization format on the encoder. Reported with an older comfy-kitchen (#15400). Update ComfyUI and its requirements.
• System RAM runs out. On #16705, one user traced this to the Video Helper Suite setting for animated previews during sampling.
• The PC freezes or the GPU is lost. A Windows RTX 5070 Ti user on #15480 stopped this by setting device to cpu on CLIPLoader, at about two minutes extra per prompt. Another on #16705 reports that --disable-cuda-graphs ended their crashes.
• aimdo memory compile error: could not start recording. --disable-comfy-compiler worked on #16342. On Strix Halo, an out-of-memory error at the first step cleared with --disable-dynamic-vram (#17063).
• Black video with NaN warnings. A report on #15524 says the fp16 video VAE works where int8 does not.
• Brightness flicker every 17 frames. MiniMax-H3 #85 reports it on full checkpoints and not on pruned ones.
• Macs. The stock int8 path failed on MPS in MiniMax-H3 #56.

Is it maintained

On 9 November, MiniMax-AI/MiniMax-H3 was not archived, had 13 commits in 90 days (last push 28 October) and no tagged releases. ComfyUI’s latest release was v0.44.0 on 4 November; update before reporting a bug.

FAQ

Can I use MiniMax H3 in the United States?

The community licence grants no rights there, or in the EU, the UK or South Korea. MiniMax invites people in those places to contact it about a licence.

What is the smallest MiniMax H3 download for ComfyUI?

35.09 GB from Comfy-Org, using the pruned_w6a8 transformer. The stock template is 42.03 GB with its Turbo LoRA.

How much VRAM does MiniMax H3 need?

ComfyUI and MiniMax give no number. The diffusers docs describe 24 to 32 GB cards with streaming, or 12 to 16 GB at a smaller canvas, with the weights in system RAM.

Can I sell videos made with MiniMax H3?

In a licensed territory, the text has no non-commercial clause and claims no rights in outputs. Above US$20M in yearly revenue you need written authorisation, and displaying the videos in an excluded territory falls outside the grant.

Related guides and tools

• Local AI video generation in ComfyUI (Wan 2.2, HunyuanVideo 1.5, LTX-2.5)
• ComfyUI install in 2026: Desktop, Portable or Manual
• ComfyUI templates: which run locally and which need the cloud
• MOSS-Image in ComfyUI
• Flux LoRA licences
• ComfyUI models and VRAM
• Model licence checker

Sources

• MiniMaxAI/MiniMax-H3 on Hugging Face (card, API parameter count, file listing) and its LICENSE
• Comfy-Org/MiniMax-H3 (README and file sizes) and FastVideo/FastVideo-FastH3-Comfy
• ComfyUI docs: H3 overview, native workflows, FastH3
• Comfy-Org/workflow_templates (index.json and the template JSON files) and the ComfyUI source (comfy/cli_args.py, comfy/supported_models.py, comfy_extras/nodes_minimax_h3.py, comfy_api_nodes/nodes_minimax.py)
• Comfy: MiniMax commercial licence and MiniMax H3 licence request form
• diffusers: MiniMax-H3 and SGLang cookbook
• ModelTC/Minimax-H3-Turbo (LoRA training specs) and Kijai/MiniMax-H3-experimental
• Comfy: Dynamic VRAM announcement
• HunyuanVideo 1.5 licence (territory comparison)

Updated 9 November 2026. File sizes, parameter counts, template contents, licence text, GitHub issues and repository activity were read on that date.