MiniMax H3 License Acknowledgement
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This checkpoint is a derivative of MiniMax H3 and is distributed under the MiniMax H3 Community License Agreement. By requesting access, you acknowledge that your use, modifications, and redistribution are subject to the original license, including its territorial and acceptable-use restrictions.
Log in or Sign Up to review the conditions and access this model content.
MiniMax H3 — Pruned INT8 ConvRot + Image Training Adapter (Fused)
Unofficial experimental derivative; not affiliated with or endorsed by MiniMax, Comfy-Org, or CircleStone Labs.
This project combines the MiniMax H3 FL2VA pruned INT8 ConvRot diffusion checkpoint from Comfy-Org with CircleStone Labs' MiniMax H3 Image Training Adapter, with the adapter fused into the model weights.
Its intended purpose is to simplify setup for image-focused LoRA training by avoiding a separate adapter fusion step at training time. It is not a newly trained base model.
Sources and attribution
- Original model: MiniMaxAI/MiniMax-H3
- Base checkpoint: Comfy-Org/MiniMax-H3, file
diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors - Image-training adapter: circlestone-labs/MiniMax-H3-Image-Training-Adapter
What was changed
The CircleStone Labs image-training adapter was merged into Comfy-Org's preexisting pruned INT8 ConvRot FL2VA checkpoint to create a single derivative .safetensors checkpoint. The base model and adapter remain the work of their respective creators; this repository documents only the fusion.
Status
Fused model weights are available through gated access.
- Fusion completed.
- Preliminary file/reload sanity checks passed.
- End-to-end image-training results have been validated.
- MiniMax has confirmed authorization to use MiniMax H3 under its Community License Agreement, subject to continued compliance.
These statements reflect preliminary preparation checks, not a guarantee of training correctness or performance.
Intended use and caveats
The fused checkpoint is intended for experimentation with image-only LoRA training in workflows that support the corresponding MiniMax H3 FL2VA checkpoint format.
- Do not apply the CircleStone adapter again when training with a checkpoint that already contains its weights.
- As the adapter author explains, image training with this adapter should use ordinary training rather than CFG-augmented/guidance-preservation training.
- The adapter author recommends using the regular base model, without the image-training adapter, for inference.
- Compatibility with a specific training framework or configuration should be verified before use; this is not a plug-and-play general inference release.
See the upstream adapter's usage notes.
License and distribution
MiniMax H3 and derivative works are subject to the MiniMax H3 Community License Agreement, including its territorial and acceptable-use restrictions. In particular, the standard agreement excludes the United States, European Union, United Kingdom, and Republic of Korea from its Applicable Territory.
The upstream LICENSE and required NOTICE files are included in this repository. Downstream recipients are responsible for ensuring that their own use, modification, and redistribution comply with the original agreement. No additional rights or geographic permissions are implied by this project or by a Hugging Face repository's visibility setting.
- Downloads last month
- 4