{
 "format": "jangtq2",
 "format_version": 2,
 "jangtq": {
  "version": 2,
  "packing": "lsb-bitstream",
  "scale_dtype": "float16",
  "rotation": "hadamard32",
  "codebook_family": "odd-cubic",
  "codebooks": {
   "2": {
    "alpha": 0.8929999999999999,
    "beta": 0.05065,
    "levels": [
     -1.5104438066482544,
     -0.4528312385082245,
     0.4528312385082245,
     1.5104438066482544
    ]
   },
   "3": {
    "alpha": 0.47124999999999995,
    "beta": 0.011599999999999997,
    "levels": [
     -2.1467249393463135,
     -1.359375,
     -0.746025025844574,
     -0.23707500100135803,
     0.23707500100135803,
     0.746025025844574,
     1.359375,
     2.1467249393463135
    ]
   },
   "4": {
    "alpha": 0.2405,
    "beta": 0.002100000000000002,
    "levels": [
     -2.689687490463257,
     -2.1399624347686768,
     -1.6721374988555908,
     -1.2736124992370605,
     -0.9317874908447266,
     -0.6340625286102295,
     -0.36783748865127563,
     -0.12051250040531158,
     0.12051250040531158,
     0.36783748865127563,
     0.6340625286102295,
     0.9317874908447266,
     1.2736124992370605,
     1.6721374988555908,
     2.1399624347686768,
     2.689687490463257
    ]
   }
  },
  "method": {
   "experts": "JANGTQ v2 odd-cubic codebook, per-row fp16 scale, rotation hadamard32",
   "rounding": "GPTQ (TQ codebook) in the rotated basis",
   "gptq_prior": "per-expert imatrix",
   "gate_up_method": [
    "gptq"
   ],
   "down_method": [
    "gptq"
   ],
   "down_gptq_layers": "42/42 (each validated on held-out rows, else RTN)",
   "allocation": "heap-MCKP over measured per-layer unit curves",
   "awq": "evaluated, not applied"
  }
 },
 "source": {
  "repo": "zai-org/GLM-5.3-Flash-BF16",
  "revision": "a5b45eb41df6402735dedc900be14a42e8d5e538"
 },
 "mtp": "dropped",
 "vision": "bf16",
 "video": "supported (shared vision tower)",
 "calibration": {
  "reference": "FP8 release",
  "tokens": 600064,
  "mix": "web50/code25/chat15/math10",
  "rotation": "hadamard32",
  "reference_captures": [
   "FP8 release: 600,064 tokens web50/code25/chat15/math10",
   "bf16 layer-streamed agentic capture: 448 GLM-template tool conversations, 171k tokens, held-out tools/sequences excluded"
  ],
  "capture_pool_weight_agentic": 2.0,
  "imatrix": "per-expert (GPTQ prior)",
  "awq": "evaluated, not applied",
  "gptq": {
   "prior": "per-expert imatrix",
   "basis": "hadamard32-rotated",
   "gate_up": "['gptq']",
   "down": "42/42 (each validated on held-out rows, else RTN)"
  },
  "build_started": "2026-09-26 02:53:30",
  "build_finished": "2026-09-26 04:25:30"
 },
 "plan": {
  "total_gib": 95.89287590235472,
  "fixed_gib_by_format": {
   "keep": 0.3928511068224907,
   "affine8": 4.035087585449219,
   "mxfp8": 4.44927978515625,
   "vision_bf16": 1.0498371124267578
  },
  "experts_gib": 85.9658203125,
  "expert_bits_histogram": {
   "gate_up:2": 36,
   "down:3": 41,
   "gate_up:3": 6,
   "down:2": 1
  }
 },
 "expert_bits": {
  "3:gate_up": 2,
  "3:down": 3,
  "4:gate_up": 2,
  "4:down": 3,
  "5:gate_up": 3,
  "5:down": 3,
  "6:gate_up": 3,
  "6:down": 3,
  "7:gate_up": 2,
  "7:down": 3,
  "8:gate_up": 2,
  "8:down": 3,
  "9:gate_up": 2,
  "9:down": 3,
  "10:gate_up": 2,
  "10:down": 3,
  "11:gate_up": 2,
  "11:down": 3,
  "12:gate_up": 2,
  "12:down": 3,
  "13:gate_up": 2,
  "13:down": 3,
  "14:gate_up": 2,
  "14:down": 3,
  "15:gate_up": 2,
  "15:down": 3,
  "16:gate_up": 2,
  "16:down": 3,
  "17:gate_up": 2,
  "17:down": 3,
  "18:gate_up": 2,
  "18:down": 3,
  "19:gate_up": 2,
  "19:down": 3,
  "20:gate_up": 2,
  "20:down": 3,
  "21:gate_up": 2,
  "21:down": 3,
  "22:gate_up": 2,
  "22:down": 3,
  "23:gate_up": 2,
  "23:down": 3,
  "24:gate_up": 2,
  "24:down": 3,
  "25:gate_up": 2,
  "25:down": 2,
  "26:gate_up": 2,
  "26:down": 3,
  "27:gate_up": 2,
  "27:down": 3,
  "28:gate_up": 2,
  "28:down": 3,
  "29:gate_up": 2,
  "29:down": 3,
  "30:gate_up": 2,
  "30:down": 3,
  "31:gate_up": 2,
  "31:down": 3,
  "32:gate_up": 3,
  "32:down": 3,
  "33:gate_up": 2,
  "33:down": 3,
  "34:gate_up": 3,
  "34:down": 3,
  "35:gate_up": 3,
  "35:down": 3,
  "36:gate_up": 3,
  "36:down": 3,
  "37:gate_up": 2,
  "37:down": 3,
  "38:gate_up": 2,
  "38:down": 3,
  "39:gate_up": 2,
  "39:down": 3,
  "40:gate_up": 2,
  "40:down": 3,
  "41:gate_up": 2,
  "41:down": 3,
  "42:gate_up": 2,
  "42:down": 3,
  "43:gate_up": 2,
  "43:down": 3,
  "44:gate_up": 2,
  "44:down": 3
 },
 "chat": {
  "sampling_defaults": {
   "temperature": 1.0,
   "top_p": 0.95
  },
  "default_sampling_mode": "thinking",
  "stop_token_ids": [
   154820,
   154827,
   154829
  ],
  "reasoning_efforts": [
   "low",
   "high",
   "max"
  ],
  "reasoning_effort_default": "max"
 },
 "capabilities": {
  "has_vision": true,
  "has_video": true,
  "has_audio": false,
  "supports_thinking": true,
  "default_reasoning": "on",
  "think_in_template": true,
  "reasoning_parser": "glm_think_block",
  "reasoning_prefill_open_tag": true,
  "tool_parser": "glm_xml_args",
  "tool_response_role": "observation",
  "family": "glm5_next"
 },
 "scale_correction": {
  "revision": 2,
  "rule": "row scale *= (<w,w>/<w,q>)^pow, clipped to [0.5, 2.0]",
  "pow": 1.0,
  "projections": [
   "gate",
   "up",
   "down"
  ],
  "reason": "MSE-optimal scales attenuate deep features; close-reasoning decision restored",
  "codes_changed": false
 }
}