MiniMax H3 Open Weights Download: FL2VA and Ref2VA for ComfyUI
Download MiniMax H3 open weights for ComfyUI: FL2VA and Ref2VA checkpoints, bf16, INT8, pruned and NVFP4 files, plus six official workflow templates.
MiniMax officially open-sourced H3, the omni-modal video model behind the Hailuo line, on August 3. The weights are published on Hugging Face, and native ComfyUI support landed the same day: pull request Comfy-Org/ComfyUI #15224 merged on August 3, adding joint audio-video generation with four new nodes and six official workflow templates.
H3 launched on July 31 (read the launch story). The open-weights release brings the model to local GPUs with multiple quantization options.
An example input image from MiniMax's official H3 multimodal context demo
H3-generated video combining a reference clip, an image, and an audio track in a single prompt (source: MiniMax blog)
System overview
H3 is a general-purpose omni-modal generation system built from three modules:
- H3-Context-IR: a managed preprocessing and orchestration system that parses the multimodal context (text, images, video, audio) and converts it into a structured intermediate representation for generation. It is not included in the open-source release; MiniMax provides it as an API and publishes a prompt-writing guide for building your own.
- H3-Base: the open-weight 33.1B omni transformer that generates 768p video with native stereo audio.
- H3-Regenerate-2K: regenerates the 768p output at 2K resolution in-context, reusing the original multimodal context. This module is not open-sourced yet; an API reproduces the full 2K pipeline.
The open release covers H3-Base as two task-specific checkpoints: FL2VA (text-to-video, first-frame, last-frame, and first+last-frame modes) and Ref2VA (reference-based generation from images, video, and audio).
Output specs
| Item | Spec |
|---|---|
| Duration | 4-15 seconds |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 |
| Resolution | Short edge 768px by default; 2K via H3-Regenerate-2K |
| Frame rate | 24 FPS |
| Audio | 32 kHz stereo, generated in the same pass |
| Languages | 11 stably supported (ar, zh, en, fr, de, it, ja, ko, pt, ru, es) |
Input modes
FL2VA accepts 0, 1, or 2 images: no image is text-to-video, one image is first- or last-frame conditioning, and two images are first+last-frame conditioning.
Ref2VA accepts up to 9 reference images, 3 reference videos (2-15 s each, 15 s total max), and 3 reference audio clips (2-15 s each, 15 s total max, always alongside an image or video). Mixed input is capped at 12 files total.
Open weights on Hugging Face
Two repositories are now live:
- MiniMaxAI/MiniMax-H3 (Hugging Face): the official release in Diffusers format, with FL2VA and Ref2VA variants, full model code, and the prompt-writing guides (base / reference). Licensed under the MiniMax H3 Community License Agreement.
- Comfy-Org/MiniMax-H3 (Hugging Face): files repackaged for ComfyUI's native nodes, ready to drop into
models/diffusion_models,models/text_encoders, andmodels/vae.
| Component | Files |
|---|---|
| Diffusion model | minimax_h3_fl2va_bf16 (61.7 GB), minimax_h3_fl2va_int8_convrot (31.7 GB), minimax_h3_fl2va_pruned_int8_convrot (19.5 GB), plus matching ref2va_* variants |
| Text encoder | qwen3vl_32b_minimax_h3_bf16 (48.0 GB), int8_convrot (25.3 GB), nvfp4_awq (14.6 GB) |
| VAEs | minimax_h3_video_vae_fp16 (4.9 GB), minimax_h3_audio_vae_fp32 (0.6 GB) |
The pruned INT8 checkpoints are about 40% smaller than the standard INT8 files thanks to precomputed adaLN curve tables, and the NVFP4 AWQ text encoder runs on any GPU. H3-Base deploys through SGLang, vLLM, diffusers, and ComfyUI.
Native ComfyUI support merged
Native support merged as Comfy-Org/ComfyUI #15224 on August 3, integrating a single-stream packed-token diffusion transformer that denoises video and stereo audio latents jointly, conditioned on Qwen3-VL-32B hidden states with per-token modality tags. Four new nodes handle the workflow: EmptyMiniMaxH3LatentAV, MiniMaxH3ImageToVideo, MiniMaxH3ReferenceToVideo, and MiniMaxH3SigmaShift.
The official text-to-video and image-to-video workflows are available as downloads:
- video_minimax_h3_t2v.json (text to video)
- video_minimax_h3_i2v.json (image to video)
- video_minimax_h3_r2v.json (reference to video)
Official workflow templates
Six official templates shipped through Comfy-Org/workflow_templates:
| Template | Mode |
|---|---|
| video_minimax_h3_t2v.json | Local: text to video |
| video_minimax_h3_i2v.json | Local: image to video |
| video_minimax_h3_r2v.json | Local: reference to video |
| api_minimax_h3_t2v.json | API: text to video |
| api_minimax_h3_r2v.json | API: reference to video |
| api_minimax_h3_flf2v.json | API: first/last frame to video |
Step-by-step guides for the local open-weights workflows are published on the official docs: MiniMax H3 open-source workflow tutorial.
Availability
- Open weights: Live on Hugging Face (MiniMaxAI/MiniMax-H3) and repackaged for ComfyUI at Comfy-Org/MiniMax-H3.
- ComfyUI: Native support merged in Comfy-Org/ComfyUI #15224; update ComfyUI to the latest version to get the new nodes.
- Official templates: Six workflows in ComfyUI's built-in template gallery.
Comments
Sign in with GitHub to join the discussion.