quantization_metadata.json

{
  "model_type": "HunyuanImage-3.0-Instruct-Distil",
  "quantization_method": "bitsandbytes_int8",
  "load_in_8bit": true,
  "llm_int8_threshold": 6.0,
  "expected_vram_gb": 18,
  "expected_total_memory_gb": 20,
  "total_params": "80B",
  "active_params": "13B (MoE)",
  "modules_kept_bf16": [
    "vae",
    "vision_model",
    "vision_aligner",
    "patch_embed",
    "final_layer",
    "time_embed",
    "time_embed_2",
    "timestep_emb",
    "guidance_emb",
    "timestep_r_emb",
    "attention_projections"
  ],
  "distil_features": {
    "cfg_distilled": true,
    "meanflow": true,
    "description": "Single-step CFG-free generation with meanflow"
  },
  "notes": "Distilled Instruct model - fast inference, single GPU friendly.",
  "attention_layers_quantized": false,
  "quality_vs_nf4": "Better quality than NF4 with reasonable memory"
}