musubi-tuner Merges Native BF16 Training Support for MiniMax-H3
bdsqlsz · x · 2026-08-06
The open-source video fine-tuning project musubi-tuner has merged a PR adding native BF16 training and inference support for the MiniMax-H3 R1 model.
Key updates include:
- T2VA, FL2VA, and JSONL-only Ref2VA latent caching.
- Qwen3-VL-32B multimodal text caching with per-row modality tags.
- BF16-base LoRA training with joint video/audio flow loss.
- H2D-only block swap to optimize VRAM usage.
Additionally, community developers are experimenting with RTX 5090 GPUs, modifying models with nvfp4 and int8convrot quantization to overcome VRAM limitations during native BF16 training.
More from Multimodal
- Hands-on: FLUX 3 in ComfyUI and MiniMax H3 on Consumer GPUs — RobbaW · 2026-08-06
- AI Creation: Generating 'The South Park Experience' with Sora 2 — RainbowUnicorns · 2026-08-06
- New AI Video Workflow Enables Perfect Character Replacement, Outperforming Kling 3.0 — breakallshittyhabits · 2026-08-06
- MiniMax H3 Tops Video Generation Categories on DesignArena — airesearch12 · 2026-08-06
- MiniMax H3 Turbo LoRA Now ComfyUI-Compatible with Tested Settings — Organix33 · 2026-08-06
- Hailuo AI Tested: Accurately Generates Videos with Indian Regional Accents — CurieuxExplorer · 2026-08-06