MiniMax H3 in Practice: Native Audio-Video Generation and Advanced Prompting

Smyshnikof · reddit · 2026-08-06

A developer shared their hands-on experience and prompting engineering details after porting the GalaxyAce LoRA to the MiniMax H3 model.

Model Features:

H3's standout feature is generating video and audio natively in a single pass (room tone, traffic, speaking), similar to the Sora 2 experience, with open weights.
The 33B model ships prequantized at around 43GB, requiring massive VRAM. Running it realistically requires an RTX 5090 (32GB) or rented Blackwell GPUs.

Prompting Techniques:

Effective Structure: Scene => Beats with timecodes => Camera behavior => Audio => No text/logos.
Pitfalls: Avoid words like "cinematic," "film grain," or "depth of field" (which imply expensive lenses), as the model tries to simulate a cheap camera sensor.
Provided a full copy-paste prompt example, noting that spoken dialogue can easily be swapped into other languages.

Training Details: Used Ostris AI-Toolkit, rank 32, taking roughly 1 hour.

More from Multimodal