quantization_config.json

{
  "quant_method": "littlebit",
  "bits_per_weight": 0.8,
  "distribution": "asymmetric_65_35",
  "asymmetric_distribution": [
    0.65,
    0.35
  ],
  "latent_factorization": "dual_svid",
  "multi_scale_compensation": true,
  "packed_bitstream": true,
  "modalities": [
    "text"
  ],
  "memory_reduction_pct": 81.7,
  "target_ram_mb": 274.5,
  "base_model": "google/embeddinggemma-2",
  "base_dimensions": 768,
  "streaming_avx2_supported": true,
  "zero_ram_inflation": true,
  "research_organization": "Samsung Research",
  "paper_title": "LittleBit: Ultra Low-Bit Quantization via Latent Factorization",
  "paper_arxiv_id": "2506.13771",
  "paper_url": "https://arxiv.org/abs/2506.13771",
  "paper_html_url": "https://arxiv.org/html/2506.13771v5"
}