MiniMax-H3-Image-MLX-Q4 / q4_components.json
zimageapp's picture
Publish original six-step H3 Image Q4 with text encoder and Image VAE
293d15e verified
Raw History Blame Contribute Delete
2.89 kB
{
"format_version": 1,
"repository": "zimageapp/MiniMax-H3-Image-MLX-Q4",
"base_model": "MiniMaxAI/MiniMax-H3-Image",
"pipeline": "text-to-image and image editing",
"quantization": {
"bits": 4,
"group_size": 64,
"mode": "affine",
"packing": "uint32, eight low-nibble-first values per word",
"dequantization": "weight = q * scale + bias",
"scale_dtype": "bfloat16",
"unquantized_tensors": "Mixed original BF16/FP32; Image VAE FP16"
},
"compatibility": {
"tensor_layout": "Native H3/Locraft naming; MLX-compatible affine packed matrices",
"validated_runtime": "Locraft C++/Metal native H3 Image",
"mlx_end_to_end": "Not implemented/validated; not a drop-in mlx-lm or Diffusers checkpoint"
},
"sampling": {
"evaluations": 6,
"sigmas": [
1.0,
0.9675236940383911,
0.8875485062599182,
0.75,
0.5327739119529724,
0.23200887441635132,
0.0
],
"guidance_scale": 1,
"loras": []
},
"components": [
{
"path": "diffusion_models/minimax_h3_image_int4.safetensors",
"bytes": 11390105242,
"sha256": "ec8e688a5918c0783b632ac50a4ac8d0ebf128d53fc29a46eb356898f8fb7c90",
"tensor_count": 996,
"source_url": "https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/0bb149f51e3156ed4b90dd45d5620346bfe1df8f/diffusion_models/minimax_h3_image_pruned_bf16.safetensors",
"source_sha256": "a767f5325db019588b3491a93d80288d336c826a7aeb87bec67675e554367222",
"modifications": "209 matrices quantized to affine Q4/group64; FP32 time/projection/output-head islands retained."
},
{
"path": "text_encoders/qwen3vl_32b_minimax_h3_int4.safetensors",
"bytes": 16466386400,
"sha256": "acb9a9dbbc8977250fe8e6aa25eb69a8294f1015332b0c6182deb31c3a9c7cd8",
"tensor_count": 1602,
"source_repository": "Comfy-Org/MiniMax-H3",
"source_family": "MiniMaxAI/MiniMax-H3",
"source_layout": "Locally quantized and merged H3 Qwen3-VL32B encoder, with embedded configuration/tokenizer.",
"modifications": "Affine Q4/group64, merged native layout, embedded tokenizer. Publication copy removes 864 bytes of inter-tensor padding; all 1602 tensor payloads and metadata verified unchanged.",
"original_aligned_file_sha256": "9a059b9dc83307f3caed3af31f4a33c78b4147194bdc3dc5adeb797caea1a25d",
"original_aligned_file_bytes": 16466387264
},
{
"path": "vae/minimax_h3_image_vae_fp16.safetensors",
"bytes": 1168505312,
"sha256": "d80e9c2e5cfc9844869961ff5a58c9a9e52182df6bfb30b26018dd7f348721af",
"tensor_count": 416,
"source_url": "https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/0bb149f51e3156ed4b90dd45d5620346bfe1df8f/vae/minimax_h3_image_vae_fp16.safetensors",
"modifications": "None; original FP16 Image VAE."
}
],
"total_weight_bytes": 29024996954
}