MiniMax H3 — Local AI Video with Native Audio
MiniMax H3 is the first open weights model in MiniMax's Hailuo video line, and it's a big one — a general purpose, omni-modal generator that understands text, images, video, and audio together and produces video with native stereo audio in a single pass, up to 2K resolution, 24fps, and around 15 seconds a clip. Voice, sound effects, and music are modeled jointly with the video, not layered on afterward. It was open sourced on August 3, 2026, with native ComfyUI support merged the same day.
✓ 100% free — download the workflow and run it locally.
Requirements
- ComfyUI (portable or desktop), updated to a recent build
- 7 custom node(s), listed below
- NVIDIA GPU recommended; see the notes for model-specific VRAM guidance
Model downloads
- minimax_h3_fl2va_pruned_int8_convrot.safetensors https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
- minimax_h3_video_vae_fp16.safetensors https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors
- minimax_h3_audio_vae_fp32.safetensors https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors
- MiniMax H3 INT4 ConvRot (text encoder only) https://huggingface.co/Merserk/MiniMax-H3-INT4-ConvRot/tree/main
Custom nodes
- ComfyUI Manager https://github.com/ltdrdata/ComfyUI-Manager
- ComfyUI-KJNodes https://github.com/kijai/ComfyUI-KJNodes
- rgthree-comfy https://github.com/rgthree/rgthree-comfy
- ComfyUI Memory Cleanup (the 🎈VRAM/RAM cleanup nodes) https://github.com/ShmuelRonen/ComfyUI-Memory-Cleanup
- ComfyUI Nvidia RTX Nodes (RTX Video Super Resolution) https://github.com/xmarre/ComfyUI-Nvidia-RTX-Nodes
- FearnworksNodes https://github.com/fearnworks/ComfyUI_FearnworksNodes
- (Optional) Spectrum MiniMax H3 - up to ~30% speed gain https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3
Setup steps
- Download the 5 model file(s) listed below and place each one into ComfyUI/models/text_encoders, ComfyUI/models/unet, ComfyUI/models/vae.
- Install the 7 custom node(s) listed below via ComfyUI Manager, then fully restart ComfyUI.
- Download the workflow .json from the button above and drag it onto the ComfyUI canvas.
- Open each loader node and confirm the model files are selected — a fresh install often defaults to the wrong one.
- Run the workflow. If a node shows red, update ComfyUI and restart before troubleshooting anything else.
Notes & tips
- Load the provided workflow into ComfyUI.
- Verify all four models are mapped correctly in the loader — the INT8 diffusion model, the NVFP4 text encoder, and both VAEs.
- For text to video, leave the first and last frame inputs empty and just write your prompt. For first/last frame control, connect a first frame, a last frame, or both.
- Write your prompt as one block that describes the shot, the motion, and the audio — dialogue, sound effects, and music all go in the same prompt, since H3 generates them together.
- Use the Resolution Selector rather than typing raw dimensions. H3's native canvas is a 768px short edge, capped at 768x1344, rounded to a multiple of 32. Keep that multiple-of-32 grid or you'll get errors.
- H3 has a minimum resolution of 384p — 256p fails completely. Start at the template's 0.5 MP preview, confirm your prompt, then raise resolution for the final render.
- Duration snaps to H3's frame grid automatically (it works in 17-frame blocks at 24fps), so don't be surprised if your exact second count rounds slightly.
- Check your system RAM, not just VRAM. During generation, RAM usage sits around 50 GB. On 16 GB or even 32 GB of RAM, the model offloads into swap on your SSD, which is slow and hammers your drive.
- If you hit an OOM on VAE Decode during longer videos, add a 🎈VRAM-Cleanup node before the VAE Decode Audio step. That alone resolved the allocation error in testing.
- On 32 GB of RAM, use a combination of 🎈VRAM-Cleanup + 🎈RAM-Cleanup to bring RAM usage down to around 30 GB. The tradeoff is that the text encoder reloads each run, but that's reads instead of writes on your SSD, and load times stay reasonable. Make sure you unset the options on those nodes exactly as shown in the workflow.
- Use SageAttention for a 20-30% speed bump. Load the model through the "Diffusion Model Loader KJ" node and select sage auto, or add the "Patch Sage Attention KJ" node after the loader.
- Launch flags that help. Starting ComfyUI with set OPTIMIZE_FOR_SPEED=1 and set PYTORCH_ALLOC_CONF=expandable_segments:True, then running with --fast, gave the best results. Add --fast-disk to load models straight from disk into VRAM and offload only what doesn't fit in RAM (watch for heavy disk writes). If you still hit OOM, add --lowvram and/or --reserve-vram 0.5 (or 1.5).
- Verify your PyTorch targets CUDA 30 or newer (cu30+). CUDA 30 added native hardware support for INT8 ConvRot; older CUDA builds fall back to software emulation and run noticeably slower.
- Free up VRAM at the source. If you have an integrated GPU, plug your monitor into the motherboard's HDMI/DP port and set it as primary in BIOS to free ~300-800 MB. No iGPU? Disabling hardware acceleration in Chrome frees a bit more.
More resources
Want the 1-click version?
Skip the manual model downloads and node installs — get a ready-to-run installer, or unlock all 75+ with Local Lab Pro.