MiniMax H3 ref2va + VBVR subset

Unmodified copies of 6 files (40.77 GB), re-hosted for operational reasons only. Nothing here has been retrained, quantised, converted or edited -- byte-for-byte the upstream blobs, in the ComfyUI category layout.

Why it exists: RunPod Serverless "cached models" pulls a whole repo onto each worker, and the upstream repos hold many variants we never load. Mirroring just the files one render touches is the difference between hundreds of gigabytes per worker and 41 GB.

Contents

file size from
diffusion_models/minimax_h3_hybrid_fl2va_ref2va_b20-49-int8.safetensors 20.97 GB smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors 15.69 GB Comfy-Org/MiniMax-H3
vae/minimax_h3_video_vae_int8_convrot.safetensors 3.17 GB Kijai/MiniMax-H3-experimental
vae/minimax_h3_audio_vae_fp32.safetensors 0.61 GB Comfy-Org/MiniMax-H3
loras/minimax_h3_ref2v_lightx2v_turbo_4step_v0.1_resized_avg_rank_20_bf16.safetensors 0.31 GB Kijai/MiniMax-H3_comfy
loras/VBVR_H3_attn_only.safetensors 0.03 GB Patarapoom/H3_lora

Licence

Each file remains under the licence of its upstream repository, linked in the table above -- this mirror grants nothing further and modifies nothing. Check the source repo before any redistribution or commercial use.

No LICENSE.md is shipped here, deliberately. The upstream repositories for these files declare no licence, so there is nothing to reproduce and inventing one would misrepresent them. Terms are whatever each source repo states.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support