Fizgig adds experimental LoRA training for MiniMax H3 with 15.7 GB model files
shootthesound · reddit · 2026-08-04
Fizgig adds experimental LoRA training for MiniMax H3
Fizgig now supports training LoRAs for MiniMax H3 in a deliberately minimal, training-only release.
- It uses the same Start → Captions → Train workflow as Klein and Krea 2.
- Users select MiniMax H3 (experimental) in the Training tab.
- The setup loads the same 15.7 GB nvfp4 Qwen3-VL-32B file used by ComfyUI instead of the 51.5 GB bf16 version, though bf16 also works.
- Adaptive LR is wired in, with two presets: MiniMax H3 Defaults and MiniMax H3 Adaptive LR.
- Download links for the DiT, text encoder, and video VAE are available in Preferences.
The one major missing piece for now: sampling during training is not supported yet.
Related event: Fizgig v3.2.0 Adds Experimental LoRA Training for MiniMax H3(2 posts)→
More from coding & agent
- Paper maps 41 agent failure modes to model, harness, tools and environment — omarsar0 · 2026-08-04
- Stanford’s CS239A course on self-improving AI agents is now on YouTube — mariofilhoml · 2026-08-04
- Cline says open-weight models work better when the harness lets them verify more — cpaik · 2026-08-04
- WorkOS schedules Agent Night in San Francisco with context-graph talks and live demos — davidcrawshaw · 2026-08-04
- A curated reading list for DeltaNet, FlashKDA, vLLM serving and MoE — austinvhuang · 2026-08-04
- Yann LeCun says strong code-generation systems go beyond plain autoregressive LLMs — ylecun · 2026-08-04