MiniMax H3 in ComfyUI: Run It Locally, VRAM by GPU, Files and Licence

MiniMax H3 runs locally in ComfyUI 0.30.0 and later through built-in templates that load a 20.97 GB INT8 video model and a 15.69 GB text encoder. MiniMax publishes no minimum VRAM for ComfyUI; its self-hosting guide cites about 18 GB of peak GPU memory for an SGLang run on one RTX 4090 with layer offloading. The open model makes 768p clips of 4 to 15 seconds with stereo audio; 2K output stays on MiniMax's API, and the licence does not cover the US, EU, UK or South Korea.

MiniMax announced H3 on 31 July 2026 and published the weights on Hugging Face on 3 August, the day ComfyUI 0.30.0 shipped native support. It reads text, images, video and audio together and generates video with a synchronised stereo soundtrack of dialogue, effects and music.

This guide covers the release, files, licence, ComfyUI route, memory from 8 GB cards to servers, published speeds and the API, as checked on 4 November 2026. Memory arithmetic follows our VRAM calculator, and our own estimates are labelled. Our comparison of open video models by GPU, Mac and licence sets H3 against Wan 2.2, LTX-2.5 and HunyuanVideo 1.5, and still pictures are covered in our guide to image models that run on your own hardware.

What MiniMax H3 is: Base, Max, Turbo and the other names

MiniMax describes the full H3 system as three modules, and only the middle one is open.

FL2VA covers text-to-video with no image, first-frame or last-frame video with one, and first-and-last-frame video with two. Ref2VA takes up to 9 images, 3 video clips and 3 audio clips as references, each clip 2 to 15 seconds and at most 12 files in all. Output runs 4 to 15 seconds at 24 fps with 32 kHz stereo audio, and the card lists stable dialogue in 11 languages.

The H3-Omni-Transformer is a dense 33-billion-parameter model. About 13 billion of those parameters sit in AdaLN branches whose outputs can be precomputed, so MiniMax says inference-only deployments need not load them. Prompts are encoded with the weights of Qwen3-VL-32B, a vision-language model of the type covered in our review of locally run vision models, and the checkpoints are CFG-distilled, which is why they sample without a negative prompt.

Search results mix in other names:

• MiniMax-H3-Max, Highspeed and Lite are API models. None has downloadable weights.
• H3 Turbo usually means the LightX2V team’s distillation LoRAs, which cut sampling to 4 or 8 steps and power ComfyUI’s optional turbo mode.
• FastH3 is the FastVideo team’s 8-step DMD2-distilled checkpoint for text-to-video and first/last-frame work.
• Pika H3 Flex is a pair of six-step adapters from Pika under the H3 licence, added to Comfy-Org’s repository on 29 October 2026.

MiniMax H3 on Hugging Face: files and sizes

The official repository totals about 498 GB because it holds the checkpoints in MiniMax’s original layout and again in Diffusers format. One original checkpoint, such as FL2VA, is 144.05 GB: a 66.28 GB transformer, a 66.73 GB text encoder and 11 GB of VAEs. MiniMax’s GitHub repository adds scripts and prompt-writing skills.

Most local users take Comfy-Org’s single-file conversions instead. Its “pruned” transformers leave out the AdaLN branches: at BF16 a pruned file is 40.23 GB against 66.28 GB for the full model, about 20 billion parameters at two bytes each.

Comfy-Org recommends the INT8 ConvRot files when PyTorch runs on CUDA 13.0, FP8 scaled only otherwise, and notes that the NVFP4 encoder does not need a Blackwell GPU. The much-searched minimax_h3_fl2va_pruned_int8_convrot.safetensors is the template default for text-to-video and image-to-video, and a full template download is about 42 GB. For live previews, taeh3.safetensors is a 22.7 MB tiny autoencoder that madebyollin added to the TAEHV project on 5 August 2026; KJNodes and ComfyUI-bleh support it.

MiniMax H3 licence: territories, revenue and acceptable use

H3 ships under the MiniMax H3 Community License Agreement, dated 2 August 2026, with Nanonoble Pte. Ltd. as licensor. It is royalty-free but not an open-source licence:

• Territory. Rights apply only in the “Applicable Territory”: the world minus the European Union, the United Kingdom, the Republic of Korea and the United States. Using the model or its outputs outside it is not authorised. Organisations in those places can apply to MiniMax for a licence, and MiniMax’s licence Q&A calls the exclusion temporary.
• Revenue. Commercial products and services earning more than 20 million US dollars a year need MiniMax’s prior written authorisation.
• Attribution. Commercial products must show “MiniMax H3” prominently in their interface; redistribution needs a NOTICE file.
• Downstream duties. Neither the model nor its outputs may be used to improve another AI model. Anyone letting third parties generate with H3 must bind them to the use restrictions, test safeguards against misuse and provide a way to report violations.

Disputes go to Hong Kong courts, MiniMax claims no rights over your outputs, and the Qwen3-VL-32B encoder carries Apache 2.0.

Comfy says it is the only official reseller of MiniMax commercial licences for local use. Professional starts at $5,000 a month for about 46,250 video-seconds and up to 10 users on the distilled weights; Enterprise is quoted per customer with a 12-month minimum and adds the undistilled weights. Neither covers inference-as-a-service or model-routing platforms. ComfyUI’s H3 tutorial says any commercial use of local outputs needs such a licence, which is stricter than the licence text and Comfy’s own licence page; where they differ, the licence governs. Our checklist for open-weight licence reviews lists the questions counsel will ask, and none of this is legal advice.

Is MiniMax H3 uncensored?

MiniMax moderates text, images, video and enhanced prompts in its hosted Context-IR stage, which is not in the open release, so a local H3-Base run skips it. The licence still applies: its acceptable use policy forbids unlawful content, content that exploits or harms minors, impersonation without consent, machine-generated content published without clear disclosure, and circumventing MiniMax’s safeguards. An organisation offering H3 to customers or other outside users must operate its own safeguards and reporting channel.

How to run MiniMax H3 in ComfyUI

1. Update ComfyUI. Native support arrived in 0.30.0 on 3 August 2026, and 0.44.0 was current on 4 November. The multiframe template needs 0.34.0, Fun ControlNet Union 0.35.0 and FastH3 0.36.0.

2. Load a template. Open Template Library, choose Video and pick MiniMax H3 Text to Video, Image to Video or Reference to Video; the model-scan pop-up downloads the files. Reference-to-video needs the Ref2VA model, which is not interchangeable with FL2VA.

3. Check the folders. On an offline machine, place the files by hand:

   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

4. Set the canvas. Use 1344 × 768 for 16:9, with sides in multiples of 32. MiniMax’s guide warns that the 1.0-megapixel preset resolves to 1376 × 768, above H3-Base’s pixel-area limit, so pick 0.98 megapixels.

5. Set the length. Frames run at 24 fps and snap to a 17k+5 grid; the default 124 frames give about 5 seconds.

6. Write the prompt. Use English and one block: the scene, then timed shots such as [Shot 2] At 00:03.500, the camera cuts to…, dialogue inside <d> tags with speaker IDs, then overall_soundscape and non_diegetic_music. Skip negative prompts, because the guider runs at CFG 1. MiniMax’s base and reference prompt guides on Hugging Face give the full format.

7. Choose the steps. The default is 20. Turbo mode’s 8-step LoRA is faster but slightly softer in motion and audio, and the 4-step Ref2VA LoRA loosens adherence to references, so keep 20 to 25 steps when identity or voice must hold.

The core nodes are MiniMaxH3ImageToVideo, with optional first and last frames, MiniMaxH3ReferenceToVideo and MiniMaxH3AddGuide. Kijai’s KJNodes pack adds memory helpers, among them a MiniMax H3 Memory Efficient Sage Attention Patch, a Low VRAM Attention node that splits attention without changing the maths, and a Token Counter.

Other routes to the same weights:

• GGUF. stable-diffusion.cpp added H3 on 4 August 2026; Unsloth’s card requires --mode vid_gen and --cfg-scale 1.0 and keeps the encoder on the CPU with --backend te=cpu. In ComfyUI, GGUF files load through leejet’s fork of ComfyUI-GGUF.
• WanGP. DeepBeepMeep’s web app for low-memory GPUs runs FL2VA and Ref2VA with accelerator profiles, on NVIDIA and on AMD RDNA 2 to 4 cards.
• Diffusers. Version 0.40.0, released on 20 August 2026, added H3 as Modular Diffusers blocks.
• Servers. MiniMax’s reference path is SGLang 0.5.17 or later, still a preview; vLLM-Omni’s experimental recipe serves both checkpoints from one process and runs on AMD ROCm.
• Mac. Draw Things’ October 2026 benchmark ran H3 at 1024 × 512 and 20 steps on Macs from a 24 GB M4 Pro to a 256 GB M5 Ultra, with its quantised builds.

MiniMax H3 VRAM requirements: 8 GB to multi-GPU

The method is parameters × bits ÷ 8. The pruned transformer’s 20 billion parameters need about 40 GB at 16-bit, 20 GB at 8-bit and 10 to 11 GB at 4-bit. The text encoder only runs while the prompt is encoded and ComfyUI can swap it out before sampling, so the transformer usually decides the card. Activations come on top and grow with resolution and length: a 15-second clip has three times the frames of a 5-second one.

WanGP’s figures are its author’s claims for the October 2026 update, not independent measurements, and they keep most weights in system memory. Offloading shifts the bottleneck to system RAM and PCIe bandwidth: vLLM-Omni’s two-GPU consumer recipe asks for at least 200 GiB of free system RAM. For ComfyUI on 8 to 16 GB cards, whose default files total about 42 GB, we would treat 64 GB of system RAM as the floor; that is our estimate, not a published requirement.

At the top end, one 96 GB RTX PRO 6000 holds even the 66.28 GB full transformer once the encoder is swapped out. Our comparison of the RTX PRO 6000 with H100 and H200 servers covers data-centre options, and our guide to choosing a GPU for local AI covers consumer cards.

Reference video, video-to-video and MiniMax H3 LoRAs

In a Ref2VA prompt, inputs are tagged <Picture 1>, <Video 1> and <Audio 1>, and the prompt says which reference controls identity, style, motion, camera or voice. Reference videos are read at 24 fps, and audio references carry a voice’s timbre, not its words. MiniMax treats video-to-video as a Ref2VA input rather than a separate task: pass the source clip as a reference and describe the edit.

ComfyUI adds related tools. MiniMaxH3AddGuide anchors stills or clips at chosen frames, which ComfyUI calls conditioning rather than video-to-video. Latent noise masks regenerate selected regions or extend a clip. Alibaba PAI’s Fun ControlNet Union patch, 4.53 GB at INT8 for version 2.0, drives motion from Canny, depth, HED, MLSD or pose videos and handles masked inpainting.

For LoRA training, Ostris’ AI Toolkit lists H3 and H3 Ref2V, kohya’s musubi-tuner has experimental support with audio-aware datasets, and DiffSynth-Studio offers LoRA and full training; on 8 September 2026 it published a training adapter for the CFG-distilled base. Community add-ons include a character-swap LoRA that replaces a person in source footage with a character from a reference image, a 0.69 GB latent upscaler, and RefMods, an unfinished MIT-licensed node pack that compresses references into small latents. LoRAs and fine-tunes are Model Derivatives, so the licence terms travel with them.

How fast is MiniMax H3 locally?

MiniMax has published no ComfyUI timings. Its self-hosting guide reproduces SGLang benchmarks, and vLLM-Omni publishes its own, for a 5-second, 1344 × 768, 124-frame clip at 50 steps unless noted.

ComfyUI defaults to 20 steps rather than 50, the Turbo LoRA and FastH3 use 8, and Pika H3 Flex uses 6. Sampling time scales roughly with steps, so by arithmetic alone an 8-step run needs about a sixth of the transformer passes of a 50-step one. Encoding, decoding and offloading do not shrink, so treat that as a ceiling on the saving, not a measurement.

MiniMax H3 local vs API: pricing and trade-offs

MiniMax’s price list on 4 November 2026, per second of output video:

Regenerating a 768p clip at 2K costs $0.05 per second, Context-IR costs $0.90 per million input tokens and $3.60 per million output tokens, and input video and extra images are billed on top. A 10-second 2K clip from MiniMax-H3 therefore costs $1.30 in output fees.

• Local keeps footage, voices and prompts on your machines, has no per-second fee and allows fine-tuning. It tops out at the open model’s 768p target, needs serious hardware and is licensed only outside the four excluded territories.
• API adds 2K output and Context-IR and is available globally, which MiniMax’s Q&A attributes to its control of the serving infrastructure. Every input is uploaded to MiniMax.

MiniMax’s documented “Full 2K Workflow” pairs a local H3-Base server with the Context-IR and Regenerate-2K APIs to match API quality, but prompts, references and the 768p result still travel to MiniMax.

How VDF AI fits

VDF AI does not generate video, and it does not run, host or integrate MiniMax H3.

The licence’s duties will be familiar to anyone who has given employees a language model: control access, keep records and leave the safeguards on. VDF AI covers that ground for language models and agents on premises, with governed agents, multi-agent orchestration in VDF AI Networks, private RAG over company documents, role-based access control on every plan and audit logs.

VDF AI Router registers custom on-premises model endpoints, such as a vLLM or Ollama server, beside the cloud models your policy approves, and its air-gap mode limits routing to the local ones.