{
 "name": "0731-affine-awq-diag-imatrix-nr8-keeper-layout-gptq-base",
 "source_repository": "deepseek-ai/DeepSeek-V4-Flash-0731",
 "source_revision": "9e165c30e2704aec5d9d593cce3eebd58bbef1cb",
 "runtime": "vMLX Python",
 "format": "JANG MLX affine with folded AWQ and diagonal imatrix",
 "profile_bits": 2,
 "bookend_bits": 8,
 "bookend_group_size": 64,
 "attention_bits": 8,
 "attention_group_size": 64,
 "token_bookend_bits": 8,
 "token_bookend_group_size": 64,
 "routed_group_size": 64,
 "routed_projection_group_sizes": {
  "w1": 64,
  "w2": 32,
  "w3": 64
 },
 "routed_projection_layer_bits": {
  "w1": {
   "5": 3,
   "14": 3,
   "30": 3,
   "34": 3,
   "37": 3,
   "42": 3
  }
 },
 "routed_projection_layer_group_sizes": {},
 "awq": {
  "capture": "actual post_attention_layernorm output after hc_pre",
  "alpha": 0.25,
  "clip": [
   0.5,
   2.0
  ],
  "fold": "inverse ffn_norm; router and routed/shared w1+w3 input columns",
  "folded_control_dtype": "F32",
  "folded_control_extra_payload_bytes": 90529792
 },
 "imatrix": {
  "capture": "actual routed down_proj input after DSV4 SwiGLU",
  "statistic": "per-channel second moment",
  "alpha": 0.25,
  "clip": [
   0.5,
   2.0
  ],
  "fold": "routed w2 columns and inverse routed w3 rows",
  "codec": "stock MLX affine"
 },
 "critical_controls": "source precision",
 "drop_mtp": true,
 "target_min_bytes": 101500000000,
 "target_max_bytes": 102500000000,
 "target_display_gib": [
  94.5,
  95.5
 ],
 "header_and_index_reserve_bytes": 100000000,
 "projected_payload_bytes": 101971512412,
 "projected_payload_gib": 94.96837147697806,
 "generation_and_prompt_contract": "official 0731 Python encoder; default thinking/low; temperature=0.6 (coding default; vendor card 1.0) top_p=0.95 top_k=0; no repetition penalty",
 "description": "Fast-regime shipping base: proven keeper NR8 layout (w1 2/g64 with six 3-bit lifts, w2 2/g32, w3 2/g64 all layers) rebuilt with pool-quant and temp-0.6 stamps; receives GPTQ error-compensated routed codes."
}