baar2_manifest.json

{
  "model_name": "BAAR2-3M-Universal-SingleHead-GAU",
  "formats": [
    "model_int4_pure.baar2",
    "model_int8_pure.baar2",
    "model.safetensors"
  ],
  "quantization": "Block-128 MSE-Search INT4 / Block-128 Pure INT8",
  "compression": "Zstandard Level 19",
  "tokenizer_embedded": true,
  "architecture": "Single-Head GAU (Hidden 256, Exp 512, Attn 64, Layers 8)"
}