MiniMax H3 INT8 base + ref2va model bundle

One public model bundle for ordinary video generation and image/video/audio reference generation. Both diffusion checkpoints use pruned INT8 ConvRot weights. All model files are copied unchanged from the pinned sources

Files

Model file Bytes
diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors 20,970,379,616
text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors 15,687,142,551
vae/minimax_h3_video_vae_fp16.safetensors 5,207,808,496
vae/minimax_h3_audio_vae_fp32.safetensors 605,254,808
loras/minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors 1,956,193,000
diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors 20,970,379,616

Total model-file storage: 65,397,158,087 bytes (60.91 GiB). This is the combined cache size, not GPU VRAM usage. Every file's source revision, size and SHA-256 is recorded in bundle-manifest.json.

Memory considerations

Switching between base and reference generation needs offloading; both diffusion models must not remain resident together. This combined path uses the FP16 video VAE, which is larger than the dedicated reference profile's INT8 video VAE. 48 GB fit and generation speed have not been GPU-verified.

Provenance and license

MiniMax H3 weights remain subject to the MiniMax H3 Community License Agreement, including its applicable-territory and usage terms. The Qwen3-VL encoder uses Apache License 2.0. See NOTICE for attribution. No new training, conversion or modification of model tensors was performed for this bundle.

Downloads last month
1,800
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for cokeadrink/minimax-h3-int8-base-ref2va

Finetuned
(232)
this model