AI Toolkit Adds MiniMax H3 Text-to-Video and Image-to-Video Training
Ostris AI Toolkit now supports training MiniMax H3 T2V and I2V LoRAs, with NVFP4 quantization and VAE gradient checkpointing for consumer GPUs.
The support landed in commit 8502a84 ("Add support for MiniMax H3 T2V and I2V training"), which ships a new minimax_h3 extension with the full training pipeline: packed-token transformer, video/audio VAE, text encoder, and NVFP4 quantization utilities.
What's included
- T2V and I2V LoRA training for both H3 checkpoints (FL2VA and Ref2VA)
- NVFP4 quantization support (
nvfp4_quant.py) for reduced VRAM during training - VAE gradient checkpointing (commit
602306d) to lower memory usage - Joint audio-video handling — the H3 pipeline denoises video and stereo audio latents together, and the toolkit preserves that during training
- Image support in the data pipeline (commit
88ac27f) for I2V training pairs - Faster quantization processing (commit
d3a3f70) tuned for H3's layer structure
Model support
MiniMax H3 is now listed in AI Toolkit's supported models as MiniMaxAI/MiniMax-H3. H3 is the 33.1B omni-modal transformer behind MiniMax's Hailuo line, open-sourced on August 3 with native ComfyUI support (read the open-weights story).
Getting started
AI Toolkit runs as a GUI or CLI on consumer hardware. Install it from the ostris/ai-toolkit repo, then create a training config pointing at the H3 base model. The minimax_h3 extension is built in — no separate install needed.
Links
- AI Toolkit GitHub: github.com/ostris/ai-toolkit
- H3 Model: MiniMaxAI/MiniMax-H3
- ComfyUI repackaged: Comfy-Org/MiniMax-H3
- License: MiniMax H3 Community License (open weights)
Comments
Sign in with GitHub to join the discussion.