Authors: Michael Helcig, Earl Killian, Dan Alistarh
Organizations: Department of Computer Science, ETH Zürich, Zurich, Switzerland · Rhoa Research, CA, USA · Institute of Science and Technology Austria, Klosterneuburg, Austria
Block-scaled lookup-table (LUT) formats, which represent weights as indices into a small value set (the grid), have become standard for low-bit LLM quantization, ranging from NF4 and BOF4 to the 3-bit LUT instruction in NVIDIA Rubin. While these formats leave the grid definition flexible, existing implementations rely on fixed grids derived either from numerical data types (INT, FP) or assumed weight distributions (NF, BOF). We demonstrate that no single fixed grid is universally optimal; the best fixed grid varies with bit-width: BOFS at 2 bits, NF at 4 and 5 bits, and INT at 6 bits. We introduce SHq, a closed-form family of grids parameterized by two critical degrees of freedom: a companding strength α that interpolates level spacing from uniform to floating-point-like, and an offset c that introduces asymmetry. This offset allows the sign bit of the FP8 block scale, otherwise unused, to select between the grid and its mirror image, up to doubling the number of distinct values available per block. SHq modifies only the lookup table contents and serves as a drop-in replacement for the grid in any such format. Across five models ranging from 3B to 35B parameters, including a mixture-of-experts architecture, evaluated at 2-6 bits using RTN and GPTQ, SHq achieves the lowest KL divergence in every setting at 3 and 4 bits and in most settings at 2, 5 and 6 bits. It outperforms the best fixed grid by up to 34% and remains superior under mixed-precision configurations. At 4.5 bits per weight, the SHq format reduces the KL divergence of NVFP4 by 41-67% while improving downstream accuracy. Grid families incorporating both spacing and offset parameters perform comparably, whereas those lacking an offset exhibit clearly higher error, confirming that these two degrees of freedom drive the observed gains. We therefore propose SHq as the standard grid for block-scaled lookup-table formats.
Figures & tables
Figure 1: Fixed grid: the lookup table holds a grid derived once for an assumed weight distribution (NF, BOF). Tunable grid (this work): it holds SHq(α,c) , with (α,c) selected for each model by the calibration KL divergence. Levels at 3 bits before E4M3 rounding; SH parameters as selected for Qwen3-8B.
Figure 2: Construction of the 3-bit SH grid selected for Qwen3-8B ( α=1.75 , c=−0.1875 ; levels before E4M3 rounding). A negative block scale selects the mirror image −G , whose levels interleave with those of G : 16 values per block, against 8, 9 and 13 for SH with c=0 , INT3 and NF3.
Llama-3.2-3B-Inst.
Qwen3-8B
Llama-3.1-8B-Inst.
Qwen3.8-27B
Qwen3.6-35B-A3B
SH
BOFS
NF
INT
SH
BOFS
NF
INT
SH
BOFS
NF
INT
SH
BOFS
NF
INT
SH
BOFS
NF
INT
2 bits (2.5 bpw incl. scale)
uniform
RTN
744.1
939.8
1228.2
1209.7
520.6
561.8
601.8
618.2
643.2
819.0
1001.7
955.4
212.2
242.2
286.0
273.8
250.6
289.7
336.7
328.7
uniform
GPTQ
577.0
782.7
813.6
985.9
379.8
378.2
432.5
405.3
509.0
628.8
727.8
685.7
194.5
212.5
249.6
230.1
208.2
224.4
263.6
240.5
3 bits (3.5 bpw incl. scale)
uniform
RTN
94.45
122.7
112.2
168.1
78.31
87.23
88.60
109.6
90.82
107.0
105.2
138.0
44.44
48.28
49.83
58.06
55.40
63.74
62.45
77.77
Table 1: KL divergence (mn, lower is better) to the BF16 model on 500K held-out tokens, one format for all layers. Each width adds 0.5 bpw for the E4M3 block scale. Bold: best per model and row.
Figure 3: Average accuracy (%) over six downstream benchmarks after RTN with one format for all layers; dashed line: BF16 model. The tasks, per-task accuracies and the GPTQ results are in Appendix G .
Qwen3-8B
Llama-3.1-8B-Inst.
SH
BOFS
NF
INT
SH
BOFS
NF
INT
3 bits
RTN
uniform
78.31
87.23
88.60
109.6
90.82
107.0
105.2
138.0
RTN
mixed
77.79
80.26
83.84
101.1
85.12
91.09
96.23
107.1
GPTQ
uniform
61.90
66.94
67.60
81.51
83.66
98.18
94.57
119.4
GPTQ
mixed
60.49
62.77
64.25
74.46
78.18
78.39
85.62
89.53
Table 2: Held-out KL (mn) of the 8B models with one code width for all layers ( uniform ) and with per-layer widths chosen by RCO at the same average width ( mixed , one seed), on RTX A6000. Bold: best per model and row.
NVFP4
SH
BOFS
NF
INT
Llama-3.2-3B-Inst.
RTN
52.18
20.36
27.86
25.60
35.56
Llama-3.2-3B-Inst.
GPTQ
52.04
16.98
30.29
25.90
37.48
Qwen3-8B
RTN
48.70
18.35
22.82
21.61
28.89
Qwen3-8B
GPTQ
32.65
14.46
16.14
15.78
19.94
Llama-3.1-8B-Inst.
RTN
52.26
21.21
26.09
25.47
34.27
Llama-3.1-8B-Inst.
GPTQ
52.91
19.81
25.13
25.63
32.93
Table 3: 4-bit formats vs NVFP4, all at 4.5 bpw. NVFP4: E2M1 codes, unsigned E4M3 block scale without scale search, FP32 tensor scale (not counted). Bold: best per row.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Llama-3.2-3B-Inst.
Qwen3-8B
Llama-3.1-8B-Inst.
Qwen3.8-27B
Qwen3.6-35B-A3B
SH
BOFS
NF
INT
SH
BOFS
NF
INT
SH
BOFS
NF
INT
SH
BOFS
NF
INT
SH
BOFS
NF
INT
2 bits (2.5 bpw incl. scale)
uniform
RTN
70.44
68.00
63.92
64.03
75.48
74.70
73.92
73.19
73.39
70.56
67.47
67.88
84.18
83.08
81.62
81.68
82.78
81.51
80.05
80.19
uniform
GPTQ
73.96
71.27
70.47
67.76
79.22
79.16
77.70
78.30
76.56
74.38
72.54
73.28
85.09
84.29
83.01
83.62
84.46
83.76
82.44
83.12
3 bits (3.5 bpw incl. scale)
uniform
RTN
88.85
87.41
87.92
85.24
90.69
90.00
90.06
88.70
89.92
88.95
89.12
87.35
93.05
92.73
92.62
91.93
92.01
91.35
91.51
90.43
Appendix
Table 4: Expected acceptance rate (EAR, %, higher is better) to the BF16 model on 500K held-out tokens, one format for all layers. Each width adds 0.5 bpw for the E4M3 block scale. Bold: best per model and row.
GPTQ run
H200
L40S
RTX A6000
differ
H200
62.65
62.69
62.92
–
H200, lazy batch 64
61.89
61.87
61.97
41%
L40S
66.03
66.12
66.16
42%
RTX A6000
61.84
61.88
61.90
42%
Appendix
Table 5: One GPTQ configuration (Qwen3-8B, SH at 3 bits) from four runs that differ only in GPU type or lazy-batch width (rows), scored on three GPU types: KL (mn). Last column: weights that differ from the H200 run.
2 bits
3 bits
4 bits
5 bits
6 bits
Llama-3.2-3B-Inst.
α
1.9375
1.875
1.9375
1.5 ∗
2.375
c
−0.46875
−0.1875
−0.078125
−0.125
−0.09375
Qwen3-8B
α
0.625
1.75
1.625 †
1.5 †
2.25
c
−0.15625
−0.1875
−0.078125
−0.03125
−0.125
Llama-3.1-8B-Inst.
α
1.375
1.875
1.625 ‡
1.75 ‡
2.5625
c
−0.359375
−0.171875
−0.0625
−0.09375
−0.109375
Appendix
Table 6: SH parameters (α,c) per model and code width, chosen by the format scan on calibration KL. † : LUT placement factor 21/4 ; ∗ : 21/2 ; ‡ : 23/4 (otherwise 1).
RTN
GPTQ
SH
NF
BOF
INT
SH
NF
BOF
INT
Qwen3-8B
2 bits
+81.7
+1947
+1815
+3175
+48.8
+37.3
+40.7
+52.6
3 bits
+32.0
+17.7
+25.8
+23.1
+20.7
+22.3
+23.6
+27.0
4 bits
+29.3
+16.6
+12.5
+10.1
+23.2
+17.3
+11.7
+13.4
Llama-3.1-8B-Inst.
2 bits
+123
+890
+898
+1195
+102
+42.9
+69.6
+94.2
3 bits
+34.0
+20.3
+19.3
+31.0
+37.9
+17.3
+23.2
+24.4
Appendix
Table 7: Increase of the held-out KL (%) when the block scale has no sign bit, with each grid’s format re-scanned without it.
Qwen3-8B
Llama-3.1-8B-Inst.
3 bits
4 bits
3 bits
4 bits
Free width and offset
NF, free width and offset (NFOC)
+0.4
+5.0
+4.8
+4.9
Arctan compander (GC, β=1 )
+1.6
+5.5
+2.3
+6.8
Heavier-tail compander (GC, β=0.25 )
+2.6
–
+2.4
+6.4
No offset
Appendix
Table 8: Calibration KL of the best grid of each family with RTN, in % relative to SH (negative: lower than SH; –: not scanned).
Qwen3-8B
Llama-3.1-8B-Inst.
SH grid of the other model
− 0.7
− 3.7
NF, free width and offset (NFOC)
+2.3
+0.2
Arctan compander (GC, β=1 )
+3.0
+1.6
Heavier-tail compander (GC, β=0.25 )
+7.7
+1.9
NF, free width (NFO)
+10.4
+16.7
NF (placement 23/4 )
+8.9
+14.0
Appendix
Table 9: Held-out KL of the best 3-bit grid of each family after GPTQ, in % relative to SH on the same GPU (single runs).
MMLU
ARC-C
HellaSwag
PIQA
SIQA
WinoGrande
Avg
KL (mn)
EAR (%)
Qwen3-8B
BF16
74.9
55.5
57.2
76.6
51.6
68.0
63.9
–
–
RTN
SH4
74.7
55.5
56.6
76.6
51.1
68.1
63.8
18.35
95.5
NVFP4
73.6
55.5
56.3
76.4
49.9
69.8
63.6
48.70
92.5
GPTQ
SH4
74.6
54.8
56.7
76.4
51.5
68.0
63.7
14.46
96.2
NVFP4
74.1
54.9
56.4
76.4
50.8
67.0
63.3
32.65
94.3
Llama-3.1-8B-Inst.
BF16
68.3
53.9
59.8
80.1
49.8
73.7
64.3
–
–
Appendix
Table 10: Benchmark accuracy (%) of SH4 and NVFP4 at 4.5 bpw, with the KL (mn) and EAR (%) of the same weights. Avg: mean of the six tasks. Bold: higher Avg of each pair.
MMLU
ARC-C
HellaSwag
PIQA
SIQA
WinoGrande
Avg
KL (mn)
EAR (%)
BF16
74.9
55.5
57.2
76.6
51.6
68.0
63.9
–
–
3 bits (3.5 bpw incl. scale)
RTN
SH
73.7
52.9
55.5
75.9
48.8
69.1
62.6
78.31
90.7
BOFS
72.7
53.8
55.4
75.7
48.4
68.2
62.4
87.23
90.0
BOF
72.4
53.5
54.9
76.7
48.7
68.4
62.4
86.46
–
NF
73.4
53.5
55.3
75.5
49.7
67.7
62.5
88.60
90.1
Appendix
Table 11: Benchmark accuracy (%) at 2 and 3 bits on Qwen3-8B, with the KL (mn) and EAR (%) of the same weights. Bold: best per column within each block (lowest KL, highest accuracy and EAR).
MMLU
ARC-C
HellaSwag
PIQA
SIQA
WinoGrande
Avg
KL (mn)
EAR (%)
BF16
68.3
53.9
59.8
80.1
49.8
73.7
64.3
–
–
3 bits (3.5 bpw incl. scale)
RTN
SH
65.0
52.4
58.1
79.4
50.2
73.2
63.1
90.82
89.9
BOFS
64.7
51.8
57.8
78.7
49.6
73.1
62.6
107.0
89.0
BOF
64.4
51.3
58.0
78.9
50.1
73.2
62.7
109.4
–
NF
64.7
51.5
58.0
78.7
49.7
72.1
62.5
105.2
89.1
Appendix
Table 12: Benchmark accuracy (%) at 2 and 3 bits on Llama-3.1-8B-Inst., with the KL (mn) and EAR (%) of the same weights. Bold: best per column within each block (lowest KL, highest accuracy and EAR).
MMLU
ARC-C
HellaSwag
PIQA
SIQA
WinoGrande
Avg
KL (mn)
EAR (%)
BF16
82.5
54.9
63.3
82.3
54.2
73.8
68.5
–
–
3 bits (3.5 bpw incl. scale)
RTN
SH
83.3
56.4
62.6
81.9
53.2
72.8
68.3
55.40
92.0
BOFS
83.6
57.6
62.4
82.3
53.3
74.2
68.9
63.74
91.3
BOF
84.1
56.9
62.7
81.7
54.4
73.4
68.9
63.35
–
NF
83.7
56.0
62.8
81.5
52.4
73.2
68.2
62.45
91.5
Appendix
Table 13: Benchmark accuracy (%) at 2 and 3 bits on Qwen3.6-35B-A3B, with the KL (mn) and EAR (%) of the same weights. Bold: best per column within each block (lowest KL, highest accuracy and EAR).
Figure 4: RTN weight error of the 252 linear layers of Qwen3-8B at 3 bits against α ( c=−0.1875 ), in % above each layer’s lowest error over the ten values. Dashed: the α chosen by the format scan.