MiniMax-H3-FP4
Quantized video model weights for RTX 50 series
2 min read
The model is an optimized version of the main MiniMax H3 video architecture (Ref2VA transformer) converted to the NVFP4 (E2M1) W4A4 format. Quantization (weight compression to save memory) reduces the weight of the core components from 66.28 GB to 37.48 GB, allowing you to run reference video generation with an audio track on a consumer RTX 5070 graphics card with 12 GB VRAM (video memory). The solution is intended for researchers and enthusiasts with the newest NVIDIA Blackwell architecture GPUs who want to work locally with heavy video generators.
Key features
- Memory savings: Transformer weight reduction to ~37 GB thanks to NVFP4 while keeping critical layers in BF16/FP32.
- RTX 50 support: Adaptation for NVIDIA Blackwell hardware tensor cores (sm120).
- Turbo mode: Accelerated generation in 8 steps (turbo8) for quick results checking.
- AdaLN caching: Pre-computed keys for different generation steps that speed up the pipeline.
What you can do
- Image-to-video generation: Create a short video clip from a reference image using a startup script.
- Local experiments: Run a heavy video generator on a mid-range card (RTX 5070 12GB) at 512p resolution or higher with tiling.
Example
python code/h3_video_job.py --ref ref.png --cond h3_cond_ref.pt --steps 8 --turbo 8 --frames 124 --size 512 --out out.mp4Read more
DetailsBenchmarks: turbo8 (124 frames / 512²) — ~203 s; 20 steps — ~284 s.
Hardware: NVIDIA Blackwell (RTX 50 series) required; no FP4 support on 40 series and AMD.
Limitations: Does not contain the text encoder (Qwen3-VL), VAE, and tokenizer — you need to grab them from the base MiniMax-H3 repository.














Comments