IAmIronMan42 · AI Infrastructure

MiniMax-H3-FineTuning

MiniMax shipped H3 weights but no trainer; this fills the gap with a two-stage supervised framework - offline latent caching of video, audio and text, then transformer optimisation under rectified-flow loss. Trains heads (~1M params), LoRA (~10.7M) or the full 33B, with DeepSpeed ZeRO-3 and a ~70k-token sequence ceiling. Practical runs want 8x A800-80GB. 304 stars.