Dataset Viewer
Auto-converted to Parquet Duplicate
gpu_name
large_stringclasses
5 values
model_base
large_stringclasses
29 values
quant
large_stringclasses
33 values
n_ctx
float64
2.05k
130k
concurrent_users
float64
1
32
backends
large_stringclasses
5 values
runner_type
large_stringclasses
5 values
mean_throughput_tok_s
float64
1.94
1.46k
std_throughput_tok_s
float64
0
280
sample_count
int64
0
22
mean_avg_ttft_ms
float64
12.9
261k
std_avg_ttft_ms
float64
0
149k
avg_ttft_ms_count
int64
0
22
mean_p50_ttft_ms
float64
13.1
279k
std_p50_ttft_ms
float64
0
159k
p50_ttft_ms_count
int64
0
22
mean_p99_ttft_ms
float64
17
584k
std_p99_ttft_ms
float64
0
753k
p99_ttft_ms_count
int64
0
22
mean_avg_itl_ms
float64
2.3
3.37k
std_avg_itl_ms
float64
0
877
avg_itl_ms_count
int64
0
22
mean_p50_itl_ms
float64
2.29
3.46k
std_p50_itl_ms
float64
0
900
p50_itl_ms_count
int64
0
22
mean_p99_itl_ms
float64
3.06
5.07k
std_p99_itl_ms
float64
0
1.01k
p99_itl_ms_count
int64
0
22
mean_avg_power_w
float64
3.2
549
std_avg_power_w
float64
0
192
avg_power_w_count
int64
0
22
mean_avg_gpu_temp_c
float64
30
82.5
std_avg_gpu_temp_c
float64
0
12.7
avg_gpu_temp_c_count
int64
0
22
ci95_low_tok_s
float64
-1,975.23
1.35k
ci95_high_tok_s
float64
2.68
3.11k
Apple M4 Pro
Qwen3.5-0.8B
BF16
4,096
null
null
tool-accuracy
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
BF16
8,192
1
Metal 4
llama-server
77.555
0.4596
2
40.955
0.0495
2
30.2
0.1556
2
106.185
0.6152
2
12.76
0.0566
2
12.75
0.0849
2
13.565
0.3889
2
null
null
0
null
null
0
73.4256
81.6844
Apple M4 Pro
Qwen3.5-0.8B
BF16
8,192
2
Metal 4
llama-server
98.36
1.3435
2
65.675
1.1809
2
56.99
0.5657
2
135.82
3.7901
2
19.965
0.2475
2
18.085
0.0071
2
27.865
0.1061
2
null
null
0
null
null
0
86.2893
110.4307
Apple M4 Pro
Qwen3.5-0.8B
BF16
8,192
4
Metal 4
llama-server
120.465
0.1344
2
107.875
8.7893
2
97.71
0.4384
2
221.535
63.0103
2
32.175
0.0778
2
32.175
0.0354
2
37.33
2.9557
2
null
null
0
null
null
0
119.2579
121.6721
Apple M4 Pro
Qwen3.5-0.8B
BF16
8,192
8
Metal 4
llama-server
121.29
1.4566
2
7,642.225
206.5671
2
8,403.35
159.2687
2
8,779.835
144.9922
2
32.375
0.5162
2
32.115
0.5728
2
54.84
0.1697
2
null
null
0
null
null
0
108.2028
134.3772
Apple M4 Pro
Qwen3.5-0.8B
BF16
8,192
16
Metal 4
llama-server
121.63
2.3052
2
22,596.71
330.6714
2
25,042.325
442.2317
2
26,174.445
432.9403
2
32.56
0.5515
2
32.14
0.5515
2
54.87
0.9192
2
null
null
0
null
null
0
100.9192
142.3408
Apple M4 Pro
Qwen3.5-0.8B
BF16
8,192
32
Metal 4
llama-server
123.215
1.1102
2
52,430.065
290.2603
2
56,977.665
1,334.5721
2
59,835.725
1,029.6677
2
32.195
0.2475
2
31.745
0.2192
2
54.57
0.9334
2
null
null
0
null
null
0
113.2408
133.1892
Apple M4 Pro
Qwen3.5-0.8B
BF16
8,192
null
null
multiturn
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
BF16
16,384
1
Metal 4
llama-server
76.82
0.198
2
41.44
0.1556
2
30.72
0.198
2
107.77
0.3111
2
12.865
0.0071
2
12.86
0
2
13.255
0.0212
2
null
null
0
null
null
0
75.0412
78.5988
Apple M4 Pro
Qwen3.5-0.8B
BF16
16,384
2
Metal 4
llama-server
101.225
4.3063
2
68.06
1.0041
2
60.83
2.1496
2
138.775
1.3364
2
19.095
0.3465
2
18.185
0.0354
2
27.75
0.0141
2
null
null
0
null
null
0
62.5352
139.9148
Apple M4 Pro
Qwen3.5-0.8B
BF16
16,384
4
Metal 4
llama-server
121.27
1.1738
2
111.275
15.0684
2
90.06
8.5418
2
291.66
162.4931
2
31.7
0.1131
2
31.8
0.0141
2
43.24
13.138
2
null
null
0
null
null
0
110.724
131.816
Apple M4 Pro
Qwen3.5-0.8B
BF16
16,384
8
Metal 4
llama-server
121.915
0.3323
2
7,468.925
13.2724
2
8,336.515
12.5511
2
8,809.285
57.212
2
32.16
0.1131
2
31.87
0.0141
2
44.51
14.6088
2
null
null
0
null
null
0
118.9291
124.9009
Apple M4 Pro
Qwen3.5-0.8B
BF16
16,384
16
Metal 4
llama-server
122.255
0.1202
2
22,053.045
703.7056
2
24,720.41
197.2686
2
25,651.605
335.6565
2
32.29
0.0424
2
31.87
0
2
44.55
14.6371
2
null
null
0
null
null
0
121.175
123.335
Apple M4 Pro
Qwen3.5-0.8B
BF16
16,384
32
Metal 4
llama-server
123.025
0.3323
2
53,312.715
152.0492
2
57,668.46
188.0197
2
60,120.575
163.1649
2
32.25
0.0566
2
31.82
0.0566
2
54.495
0.2758
2
null
null
0
null
null
0
120.0391
126.0109
Apple M4 Pro
Qwen3.5-0.8B
BF16
32,768
1
Metal 4
llama-server
77.26
0.2546
2
41.565
0.0636
2
30.635
0.1202
2
107.965
0.0212
2
12.815
0.0495
2
12.805
0.0636
2
13.16
0.099
2
null
null
0
null
null
0
74.9729
79.5471
Apple M4 Pro
Qwen3.5-0.8B
BF16
32,768
2
Metal 4
llama-server
92.85
11.0167
2
66.235
1.2657
2
56.09
0.0141
2
134.25
2.291
2
21.265
2.4395
2
22.64
6.2367
2
27.82
0.0424
2
null
null
0
null
null
0
-6.1297
191.8297
Apple M4 Pro
Qwen3.5-0.8B
BF16
32,768
4
Metal 4
llama-server
121.58
0.1838
2
100.02
2.1637
2
95.35
3.2951
2
175.56
0.5374
2
31.705
0.2616
2
31.8
0.0283
2
42.865
11.3491
2
null
null
0
null
null
0
119.9282
123.2318
Apple M4 Pro
Qwen3.5-0.8B
BF16
32,768
8
Metal 4
llama-server
122.37
0.3394
2
7,505.75
46.6549
2
8,330.89
0.1131
2
8,769.6
4.4689
2
32.14
0.0566
2
31.84
0.0141
2
53.885
0.2899
2
null
null
0
null
null
0
119.3206
125.4194
Apple M4 Pro
Qwen3.5-0.8B
BF16
32,768
16
Metal 4
llama-server
122.865
0.0495
2
22,605.92
77.9232
2
24,885.095
5.6356
2
25,955.84
41.9031
2
32.235
0.0212
2
31.865
0.0071
2
54.595
0.2192
2
null
null
0
null
null
0
122.4203
123.3097
Apple M4 Pro
Qwen3.5-0.8B
BF16
32,768
32
Metal 4
llama-server
122.645
0.1344
2
51,708.22
387.1692
2
57,695.895
245.8257
2
60,164.995
361.2538
2
32.34
0.0566
2
31.865
0.0212
2
55.71
0.9899
2
null
null
0
null
null
0
121.4379
123.8521
Apple M4 Pro
Qwen3.5-0.8B
BF16
130,064
1
Metal 4
llama-server
76.875
0.0354
2
41.875
0.0919
2
31.1
0
2
108
0.5233
2
12.875
0.0071
2
12.87
0
2
13.225
0.0212
2
null
null
0
null
null
0
76.5573
77.1927
Apple M4 Pro
Qwen3.5-0.8B
BF16
130,064
2
Metal 4
llama-server
103.225
2.2557
2
68.88
0.5374
2
61.425
1.8031
2
139.945
0.8556
2
18.865
0.1202
2
18.215
0.0071
2
27.73
0.0141
2
null
null
0
null
null
0
82.9589
123.4911
Apple M4 Pro
Qwen3.5-0.8B
BF16
130,064
4
Metal 4
llama-server
119.1
2.39
2
106.165
7.4741
2
97.19
2.0223
2
215.035
55.4725
2
31.92
0.1838
2
31.805
0.0212
2
43.305
13.1027
2
null
null
0
null
null
0
97.6269
140.5731
Apple M4 Pro
Qwen3.5-0.8B
BF16
130,064
8
Metal 4
llama-server
122.745
0.2899
2
7,561.45
170.0309
2
8,339.32
17.6211
2
8,734.595
50.2965
2
32.07
0.1414
2
31.835
0.0071
2
44.32
14.3684
2
null
null
0
null
null
0
120.1403
125.3497
Apple M4 Pro
Qwen3.5-0.8B
BF16
130,064
16
Metal 4
llama-server
122.635
0.2899
2
22,537.63
8.2166
2
24,834.555
21.4041
2
25,999.765
39.704
2
32.205
0.0636
2
31.855
0.0354
2
54.915
0.5869
2
null
null
0
null
null
0
120.0303
125.2397
Apple M4 Pro
Qwen3.5-0.8B
BF16
130,064
32
Metal 4
llama-server
123.06
0.396
2
52,369.19
480.3801
2
57,319.3
14.8775
2
60,300.87
1.0324
2
32.235
0.1202
2
31.8
0.0849
2
54.815
0.2192
2
null
null
0
null
null
0
119.5023
126.6177
Apple M4 Pro
Qwen3.5-0.8B
BF16
null
null
null
context-rot
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
4,096
null
null
tool-accuracy
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
8,192
1
Metal 4
llama-server
108.275
0.2475
2
39.245
0.4596
2
28.75
0.1414
2
108.18
0.1697
2
9.105
0.0071
2
9.08
0.0141
2
9.535
0.0636
2
null
null
0
null
null
0
106.0515
110.4986
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
8,192
2
Metal 4
llama-server
120.935
8.959
2
66.455
12.735
2
46.425
0.0354
2
214.965
120.5971
2
16.145
0.9546
2
15.51
0.0707
2
18.13
3.0123
2
null
null
0
null
null
0
40.4425
201.4275
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
8,192
4
Metal 4
llama-server
141.135
3.6133
2
99.67
2.4749
2
87.99
1.6971
2
250.615
1.2233
2
26.615
0.4172
2
26.79
0
2
29.02
0.0283
2
null
null
0
null
null
0
108.6712
173.5988
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
8,192
8
Metal 4
llama-server
145.01
0.3253
2
6,549.18
146.2155
2
7,029.7
4.681
2
7,564.77
30.4056
2
27.065
0.0495
2
26.825
0.0071
2
29.1
0.0141
2
null
null
0
null
null
0
142.0876
147.9324
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
8,192
16
Metal 4
llama-server
145.49
0.3394
2
19,249.865
300.4002
2
20,952
51.4774
2
22,183.575
33.128
2
27.21
0.0141
2
26.855
0.0495
2
45.255
0.4031
2
null
null
0
null
null
0
142.4406
148.5394
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
8,192
32
Metal 4
llama-server
145.23
0.3394
2
45,178.83
670.012
2
48,769.335
11.5329
2
50,788.495
268.3117
2
27.305
0.0495
2
26.905
0.0636
2
46.465
1.5344
2
null
null
0
null
null
0
142.1806
148.2794
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
8,192
null
null
multiturn
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
16,384
1
Metal 4
llama-server
106.925
0.0071
2
39.495
0.2616
2
29.355
0.0212
2
108.455
0.2899
2
9.23
0
2
9.23
0
2
9.705
0.0636
2
null
null
0
null
null
0
106.8615
106.9885
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
16,384
2
Metal 4
llama-server
118.51
0.4525
2
61.31
0.297
2
54.26
0.396
2
170.235
53.9027
2
16.2
0.0141
2
15.615
0.0212
2
20.505
0.0354
2
null
null
0
null
null
0
114.4441
122.5759
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
16,384
4
Metal 4
llama-server
144.86
1.5698
2
104.905
20.7677
2
79.48
5.4871
2
286.925
172.7957
2
26.935
0.0778
2
26.96
0
2
36.81
10.946
2
null
null
0
null
null
0
130.7563
158.9637
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
16,384
8
Metal 4
llama-server
144.34
0.2121
2
6,397.64
95.8695
2
7,081.67
8.9095
2
7,486.435
50.7208
2
27.35
0.0566
2
27.01
0.0141
2
46.475
2.5102
2
null
null
0
null
null
0
142.4341
146.2459
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
16,384
16
Metal 4
llama-server
144.43
0.0707
2
19,094.23
705.6077
2
21,085.135
47.6802
2
22,294.355
78.4394
2
27.375
0.0071
2
27.03
0.0141
2
47.68
1.3294
2
null
null
0
null
null
0
143.7947
145.0653
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
16,384
32
Metal 4
llama-server
144.49
0.1273
2
45,216.545
57.5938
2
48,782.44
503.955
2
51,044.27
1,197.2449
2
27.43
0.0283
2
27.03
0
2
48.225
0.3182
2
null
null
0
null
null
0
143.3465
145.6335
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
32,768
1
Metal 4
llama-server
107.535
1.1384
2
39.245
0.6859
2
28.885
0.0636
2
107.95
0.8202
2
9.16
0.1131
2
9.155
0.1344
2
9.6
0.1273
2
null
null
0
null
null
0
97.3067
117.7633
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
32,768
2
Metal 4
llama-server
112.135
9.4116
2
57.585
1.0253
2
50.83
2.3193
2
127.48
4.2144
2
16.995
0.6435
2
17.84
3.0123
2
20.67
0.0283
2
null
null
0
null
null
0
27.5766
196.6934
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
32,768
4
Metal 4
llama-server
143.725
1.0253
2
106.875
5.9609
2
90.385
1.1243
2
309.675
72.4289
2
26.48
0.0141
2
26.96
0
2
29.215
0.1344
2
null
null
0
null
null
0
134.5132
152.9368
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
32,768
8
Metal 4
llama-server
144.55
0.3394
2
6,436.2
117.2807
2
7,067.21
8.7964
2
7,559.725
77.0534
2
27.215
0.0071
2
26.99
0.0141
2
31.61
3.2385
2
null
null
0
null
null
0
141.5006
147.5994
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
32,768
16
Metal 4
llama-server
144.255
0.3182
2
19,139.45
54.2492
2
21,049.63
45.0993
2
22,181.545
39.3788
2
27.395
0.0071
2
27.015
0.0071
2
48.375
0.0071
2
null
null
0
null
null
0
141.3962
147.1138
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
32,768
32
Metal 4
llama-server
144.425
0.1485
2
44,779.945
151.4269
2
48,683.6
71.6582
2
51,066.325
477.9971
2
27.455
0.0071
2
27.025
0.0071
2
48.395
0.7425
2
null
null
0
null
null
0
143.0909
145.7591
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
130,064
1
Metal 4
llama-server
106.825
0.2899
2
39.2
0.1697
2
29.395
0.0495
2
107.38
1.5981
2
9.225
0.0071
2
9.23
0
2
9.71
0.0849
2
null
null
0
null
null
0
104.2203
109.4297
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
130,064
2
Metal 4
llama-server
119.29
4.0447
2
62.79
6.1518
2
49.85
3.677
2
169.78
55.4372
2
16.145
0.5728
2
15.665
0.0495
2
20.51
0.1273
2
null
null
0
null
null
0
82.9508
155.6292
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
130,064
4
Metal 4
llama-server
143.465
2.3264
2
110.295
14.9553
2
82.505
7.46
2
331.23
112.3876
2
26.73
0.3111
2
26.96
0.0141
2
29.135
0.1768
2
null
null
0
null
null
0
122.5636
164.3664
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
130,064
8
Metal 4
llama-server
144.585
0.1202
2
6,605.2
54.4896
2
7,077.615
6.965
2
7,573.18
66.3973
2
27.22
0.0849
2
27.01
0.0141
2
37.59
11.7097
2
null
null
0
null
null
0
143.505
145.665
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
130,064
16
Metal 4
llama-server
144.23
0.1838
2
19,407.22
378.7264
2
21,072.03
51.5905
2
22,269.68
117.1252
2
27.385
0.0071
2
27.03
0.0283
2
47.09
1.3859
2
null
null
0
null
null
0
142.5782
145.8818
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
130,064
32
Metal 4
llama-server
144.43
0.0141
2
44,530.51
213.7725
2
48,955.865
112.7199
2
50,902.11
797.6023
2
27.415
0.0212
2
27.01
0.0283
2
48.445
0.7566
2
null
null
0
null
null
0
144.3029
144.5571
Apple M4 Pro
Qwen3.5-0.8B
IQ4_NL
null
null
null
context-rot
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
4,096
null
null
tool-accuracy
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
8,192
1
Metal 4
llama-server
107.355
0.1344
2
40.06
0.099
2
29.59
0.1838
2
109.275
0.2616
2
9.18
0
2
9.16
0.0141
2
9.64
0.0283
2
null
null
0
null
null
0
106.1479
108.5621
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
8,192
2
Metal 4
llama-server
118.56
3.9457
2
59.615
2.1567
2
50.845
4.6315
2
132.15
1.7961
2
16.435
0.4172
2
15.74
0.0283
2
20.495
0.0354
2
null
null
0
null
null
0
83.1103
154.0097
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
8,192
4
Metal 4
llama-server
140.605
0.9546
2
116.295
13.329
2
77.61
2.9557
2
410.705
90.5874
2
27.625
0.1485
2
27.83
0.0283
2
29.92
0.0424
2
null
null
0
null
null
0
132.0284
149.1816
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
8,192
8
Metal 4
llama-server
140.29
0.8768
2
6,682.83
130.4046
2
7,276.045
24.8972
2
7,834.195
115.4352
2
28.03
0.3394
2
27.83
0.0566
2
38.435
11.9996
2
null
null
0
null
null
0
132.4123
148.1677
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
8,192
16
Metal 4
llama-server
140.04
0.1131
2
19,946.875
256.2201
2
21,744.32
13.0673
2
22,949.77
1.1738
2
28.255
0.0071
2
27.885
0.0071
2
46.405
0.6859
2
null
null
0
null
null
0
139.0235
141.0565
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
8,192
32
Metal 4
llama-server
140.285
0.0212
2
46,367.955
699.2367
2
50,270.71
532.8191
2
53,250.385
84.3649
2
28.27
0.0141
2
27.865
0.0071
2
50.11
0.1838
2
null
null
0
null
null
0
140.0944
140.4756
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
8,192
null
null
multiturn
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
16,384
1
Metal 4
llama-server
108.87
0.1414
2
39.675
0.4738
2
29.08
0.2121
2
107.715
0.4172
2
9.06
0.0141
2
9.075
0.0354
2
9.63
0
2
null
null
0
null
null
0
107.5994
110.1406
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
16,384
2
Metal 4
llama-server
122.63
3.6487
2
63.52
5.8124
2
51.745
4.3911
2
169.83
53.8815
2
15.735
0.3465
2
15.45
0
2
20.265
0.1626
2
null
null
0
null
null
0
89.8485
155.4115
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
16,384
4
Metal 4
llama-server
141.115
0.1768
2
108.09
16.1362
2
81.995
4.8578
2
330.81
107.9611
2
27.62
0.1273
2
27.62
0.0424
2
29.82
0.1556
2
null
null
0
null
null
0
139.5268
142.7032
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
16,384
8
Metal 4
llama-server
140.78
0.3818
2
6,609.395
13.4987
2
7,236.31
12.8269
2
7,681.59
17.8332
2
28
0.0707
2
27.67
0.0141
2
47.775
1.7324
2
null
null
0
null
null
0
137.3494
144.2106
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
16,384
16
Metal 4
llama-server
141.23
0.2121
2
19,989.53
164.6852
2
21,589.66
36.5291
2
22,651.38
179.5344
2
28.04
0.1273
2
27.665
0.0636
2
47.715
2.8921
2
null
null
0
null
null
0
139.3241
143.1359
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
16,384
32
Metal 4
llama-server
141.21
0.2828
2
45,926.61
97.906
2
50,034.71
212.6836
2
52,748.355
19.4525
2
28.1
0.0566
2
27.71
0.0424
2
47.19
2.0648
2
null
null
0
null
null
0
138.6688
143.7512
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
32,768
1
Metal 4
llama-server
108.72
0.2121
2
39.545
0.1202
2
29.175
0.1768
2
108.94
0.4525
2
9.06
0.0283
2
9.045
0.0212
2
9.66
0.0141
2
null
null
0
null
null
0
106.8141
110.6259
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
32,768
2
Metal 4
llama-server
119.1
9.4187
2
63.02
7.2408
2
50.14
0.9758
2
168.205
59.2485
2
16.215
1.0677
2
15.575
0.2333
2
20.34
0.4101
2
null
null
0
null
null
0
34.478
203.722
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
32,768
4
Metal 4
llama-server
138.765
0.2616
2
92.885
5.5508
2
81.47
2.5739
2
207.985
63.675
2
27.82
0.0849
2
27.82
0.0566
2
37.275
10.2036
2
null
null
0
null
null
0
136.4144
141.1156
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
32,768
8
Metal 4
llama-server
139.87
0.495
2
6,792.965
11.2784
2
7,287.34
1.9799
2
7,791.31
49.0308
2
28.065
0.0495
2
27.885
0.0495
2
39.555
13.1027
2
null
null
0
null
null
0
135.4229
144.3171
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
32,768
16
Metal 4
llama-server
139.87
0.0424
2
19,983.655
118.7869
2
21,772.12
6.8165
2
23,027.5
36.0624
2
28.295
0.0636
2
27.955
0.0071
2
48.235
1.2799
2
null
null
0
null
null
0
139.4888
140.2512
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
32,768
32
Metal 4
llama-server
140.35
0.6788
2
46,011.33
130.744
2
50,577.775
395.1525
2
53,216.605
22.0829
2
28.265
0.1344
2
27.875
0.1202
2
49.79
0.3253
2
null
null
0
null
null
0
134.2511
146.4489
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
130,064
1
Metal 4
llama-server
107.87
0.1273
2
39.285
0.1202
2
29.81
0.099
2
109.115
0.0778
2
9.145
0.0071
2
9.14
0.0141
2
9.61
0.0707
2
null
null
0
null
null
0
106.7265
109.0135
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
130,064
2
Metal 4
llama-server
110.1
13.435
2
59.17
2.6304
2
50.755
4.9851
2
132.41
1.0182
2
17.675
1.8455
2
17.815
3.0759
2
20.55
0.1838
2
null
null
0
null
null
0
-10.607
230.807
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
130,064
4
Metal 4
llama-server
140.5
1.3011
2
104.405
3.2598
2
83.305
6.5266
2
301.355
64.0285
2
27.705
0.0071
2
27.77
0.0141
2
29.945
0.1626
2
null
null
0
null
null
0
128.8105
152.1895
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
130,064
8
Metal 4
llama-server
140.315
0.4031
2
6,762.77
36.0907
2
7,285.465
0.7707
2
7,720.675
2.5385
2
28.115
0.0495
2
27.81
0.0283
2
46.145
0.0212
2
null
null
0
null
null
0
136.6938
143.9362
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
130,064
16
Metal 4
llama-server
140.145
0.3748
2
19,678.255
15.521
2
21,674.49
14.2128
2
22,913.5
10.0268
2
28.225
0.0071
2
27.835
0.0071
2
49.725
0.8415
2
null
null
0
null
null
0
136.7779
143.5121
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
130,064
32
Metal 4
llama-server
140.53
0.0849
2
46,275.305
116.9342
2
50,420.61
71.8986
2
52,906.43
181.9951
2
28.245
0.0071
2
27.83
0.0283
2
48.8
1.9233
2
null
null
0
null
null
0
139.7676
141.2924
Apple M4 Pro
Qwen3.5-0.8B
IQ4_XS
null
null
null
context-rot
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
4,096
null
null
tool-accuracy
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
8,192
1
Metal 4
llama-server
102.88
0.0424
2
40.44
0.4667
2
29.595
0.4596
2
111.34
0.1414
2
9.595
0.0071
2
9.585
0.0071
2
10.065
0.0071
2
null
null
0
null
null
0
102.4988
103.2612
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
8,192
2
Metal 4
llama-server
112.985
2.5385
2
58.005
0.2899
2
49.72
2.0648
2
131.5
4.5255
2
16.91
0.3253
2
16
0.0566
2
21.3
0.1131
2
null
null
0
null
null
0
90.1777
135.7923
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
8,192
4
Metal 4
llama-server
136.8
0.7354
2
100.18
4.2992
2
86.975
3.2032
2
254.535
11.9289
2
28.59
0.0849
2
28.555
0.0071
2
30.81
0.0849
2
null
null
0
null
null
0
130.1929
143.4071
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
8,192
8
Metal 4
llama-server
137.8
2.1779
2
6,855.55
139.6253
2
7,365.055
143.5073
2
7,961.5
95.2614
2
28.39
0.4101
2
28.05
0.5657
2
39.845
12.5653
2
null
null
0
null
null
0
118.2328
157.3672
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
8,192
16
Metal 4
llama-server
137.2
0.6647
2
20,607.935
201.9426
2
22,261.455
7.4176
2
23,281.135
165.8378
2
28.805
0.1061
2
28.47
0.0424
2
49.255
2.2415
2
null
null
0
null
null
0
131.2282
143.1718
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
8,192
32
Metal 4
llama-server
136.945
0.0212
2
47,912.02
226.7974
2
51,300.065
155.1746
2
53,074.735
233.2109
2
28.975
0.0071
2
28.57
0
2
49.5
1.5415
2
null
null
0
null
null
0
136.7544
137.1356
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
8,192
null
null
multiturn
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
0
null
null
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
16,384
1
Metal 4
llama-server
103.19
0.1556
2
39.82
0
2
29.685
0.1202
2
111.045
0.0495
2
9.555
0.0071
2
9.55
0
2
10.08
0.0707
2
null
null
0
null
null
0
101.7923
104.5877
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
16,384
2
Metal 4
llama-server
123.65
0.2687
2
67.545
0.3182
2
46.925
0.0636
2
213.43
1.0324
2
15.965
0.0354
2
15.965
0.0354
2
16.525
0.0636
2
null
null
0
null
null
0
121.2359
126.0641
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
16,384
4
Metal 4
llama-server
136.59
0.7354
2
114.245
7.5873
2
89.13
13.4492
2
351.25
0.8627
2
28.325
0.2616
2
28.525
0.0212
2
30.59
0.0566
2
null
null
0
null
null
0
129.9829
143.1971
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
16,384
8
Metal 4
llama-server
136.455
0.9263
2
7,011.42
22.5708
2
7,489.285
6.1165
2
7,997.485
119.7485
2
28.72
0.1131
2
28.565
0.0212
2
40.31
13.4916
2
null
null
0
null
null
0
128.1326
144.7774
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
16,384
16
Metal 4
llama-server
137
0.396
2
20,587.525
40.7081
2
22,243.44
28.1146
2
23,617.235
1,180.4511
2
28.855
0.1202
2
28.53
0.0707
2
39.3
11.936
2
null
null
0
null
null
0
133.4423
140.5577
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
16,384
32
Metal 4
llama-server
136.83
0.198
2
47,652.32
518.2668
2
50,799.495
449.43
2
52,562.21
145.0983
2
28.995
0.0495
2
28.575
0.0636
2
50.335
0.601
2
null
null
0
null
null
0
135.0512
138.6088
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
32,768
1
Metal 4
llama-server
102.845
0.0071
2
40.145
0.1202
2
29.465
0.3323
2
111.67
0.1556
2
9.59
0.0141
2
9.585
0.0071
2
10.115
0.1202
2
null
null
0
null
null
0
102.7815
102.9085
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
32,768
2
Metal 4
llama-server
122.785
1.393
2
67.86
0.7495
2
46.875
0.0919
2
212.505
2.8496
2
15.985
0.0495
2
15.965
0.0071
2
18.72
3.2244
2
null
null
0
null
null
0
110.2696
135.3004
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
32,768
4
Metal 4
llama-server
136.125
0.502
2
103.745
17.0342
2
87.755
9.355
2
258.325
127.0176
2
28.28
0.2687
2
28.425
0.0071
2
37.715
10.2743
2
null
null
0
null
null
0
131.6144
140.6356
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
32,768
8
Metal 4
llama-server
137.12
0.1697
2
6,986.87
99.0232
2
7,480.84
14.7927
2
7,970.405
105.8468
2
28.71
0.2121
2
28.495
0.0778
2
40.265
13.7108
2
null
null
0
null
null
0
135.5953
138.6447
Apple M4 Pro
Qwen3.5-0.8B
Q3_K_M
32,768
16
Metal 4
llama-server
136.755
0.2616
2
20,635.885
376.202
2
22,272.605
7.3751
2
23,183.275
120.0314
2
28.935
0.1061
2
28.56
0.0566
2
50.27
0.9051
2
null
null
0
null
null
0
134.4044
139.1056
End of preview. Expand in Data Studio

Poor Paul's Benchmark Results

Community-submitted LLM inference benchmark data from real consumer, prosumer, and small-business hardware.

Each row is one normalized benchmark result: a specific model × quantization × hardware × configuration combination with measured performance metrics.

Use this to:

  • Compare throughput, latency, and power efficiency across GPU models and quantizations
  • Study how concurrency and context length scale on different hardware
  • Build leaderboards, dashboards, and data-driven GPU purchasing decisions
  • Query via AI: any MCP client connected to mcp.poorpaul.dev can answer questions from this data directly

Files

File Description
data/results_*.jsonl Raw benchmark submissions, one file per run, appended continuously
llms.txt Machine-readable summary for LLM context injection

Quick start

import pandas as pd
from datasets import load_dataset

# Stream all rows (recommended for large queries)
ds = load_dataset("paulplee/ppb-results", streaming=True, split="train")
df = pd.DataFrame(iter(ds))

# Filter to throughput rows for a specific GPU
tput = df[
    (df["gpu_name"] == "NVIDIA GeForce RTX 5090") &
    (df["runner_type"] == "llama-bench")
][["model_base", "quant", "n_ctx", "throughput_tok_s"]]

Or via DuckDB directly from HuggingFace:

SELECT gpu_name, model_base, quant, concurrent_users,
       AVG(throughput_tok_s) AS mean_tok_s,
       COUNT(*)              AS n
FROM   read_parquet('hf://datasets/paulplee/ppb-results/data/*.jsonl')
WHERE  runner_type = 'llama-bench'
GROUP  BY ALL
ORDER  BY mean_tok_s DESC;

Schema (v0.9.0)

All columns are present on every row. Fields that do not apply to a given runner are null.

Model identity

Column Type Description
run_type string quantitative, qualitative, or all
model string Full model path (e.g. unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.gguf)
model_base string Base model name without quant suffix (e.g. Qwen3.5-9B)
quant string Quantization format (e.g. Q4_K_M, Q8_0, BF16)
model_org string|null HuggingFace organisation (e.g. unsloth); null for local paths
model_repo string|null Full HF org/repo string; null for local paths
runner_type string Benchmark backend: llama-bench, llama-server, or llama-server-loadtest

LLM engine

Column Type Description
llm_engine_name string|null Inference engine (e.g. llama.cpp)
llm_engine_version string|null Engine version with build hash (e.g. b5063 (58ab80c3))

Hardware

Column Type Description
gpu_name string|null Primary GPU model name
gpu_vram_gb float|null Primary GPU VRAM in GB
gpu_driver string|null GPU driver version
gpu_count int Number of GPUs used
gpu_names string|null Comma-joined list of all GPU names (multi-GPU runs)
gpu_total_vram_gb float|null Total VRAM across all GPUs
unified_memory bool|null true for Apple Silicon — GPU and CPU share the same memory pool
gpu_compute_capability string|null CUDA compute capability (e.g. "9.0" for Blackwell); null for non-CUDA
gpu_pcie_gen int|null PCIe generation (e.g. 5); null for unified-memory platforms
gpu_pcie_width int|null PCIe link width in lanes (e.g. 16); null for unified-memory platforms
gpu_power_limit_w float|null Configured TDP limit in Watts (from NVML); null for non-NVIDIA
backends string|null Compute backend with version (e.g. CUDA 13.0, Metal, CPU)
cpu_model string|null CPU model name

Benchmark configuration

Column Type Description
n_ctx int|null Context window size in tokens
n_batch int|null Batch size for prompt processing
split_mode string|null Multi-GPU split strategy (layer, row, none); null for single-GPU
tensor_split string|null Per-GPU VRAM weight string (e.g. "1,1"); null for single-GPU
concurrent_users int|null Number of simulated parallel users. For llama-server-loadtest, each row is one concurrency level from the measured curve.

Workload

Column Type Description
task_type string|null Workload category (e.g. text-generation, context-rot-niah)
prompt_dataset string|null Prompt source (e.g. sharegpt-v3); null for llama-bench
num_prompts int|null Prompts sent per run; null for llama-bench
n_predict int|null Max tokens generated per prompt; null for llama-bench

Performance — throughput

Column Type Description
throughput_tok_s float|null Tokens per second (primary throughput metric)
vram_cliff_tokens int|null Largest n_ctx that loaded without OOM during pre-flight discovery

Performance — power

Column Type Description
avg_power_w float|null Average GPU power draw in Watts
max_power_w float|null Peak GPU power draw in Watts

Performance — thermal

Column Type Description
avg_gpu_temp_c float|null Average GPU temperature (°C)
max_gpu_temp_c float|null Peak GPU temperature (°C)
avg_cpu_temp_c float|null Average CPU temperature (°C)
max_cpu_temp_c float|null Peak CPU temperature (°C)
avg_fan_speed_rpm float|null Average fan speed (RPM)
max_fan_speed_rpm float|null Peak fan speed (RPM)

Performance — user experience (server runners only)

Column Type Description
avg_ttft_ms float|null Average Time-To-First-Token (ms)
p50_ttft_ms float|null Median TTFT (ms)
p99_ttft_ms float|null 99th-percentile TTFT (ms)
avg_itl_ms float|null Average Inter-Token Latency (ms)
p50_itl_ms float|null Median ITL (ms)
p99_itl_ms float|null 99th-percentile ITL (ms)

Qualitative evaluation

Populated when run_type is qualitative or all. All null for pure quantitative runs.

Column Type Description
context_rot_score float|null Mean accuracy across all (length × depth) long-context recall cases
context_rot_accuracy_by_length string|null JSON {haystack_length: accuracy} map
context_rot_accuracy_by_depth string|null JSON {depth_pct: accuracy} map
tool_selection_accuracy float|null Fraction of cases with correct tool name selected
parameter_accuracy float|null Fraction of cases with all required arguments matching ground truth
parameter_hallucination_rate float|null Fraction of cases with invented arguments not in schema
parse_success_rate float|null Fraction of cases with parseable tool-call JSON
overall_tool_accuracy float|null Geometric mean of tool selection × parameter accuracy
knowledge_accuracy_mean float|null Mean fraction of factual claims judged consistent with common knowledge
knowledge_accuracy_std float|null Standard deviation of per-prompt knowledge-accuracy scores
answer_relevancy_mean float|null Mean judge-rated response relevancy (0–1)
coherence_mean float|null Mean judge-rated coherence (0–1)
quality_composite_score float|null Mean of knowledge accuracy, relevancy, and coherence
memory_accuracy float|null LongMemEval recall accuracy (0–1); null when MT-Bench mode was used
mt_bench_score float|null MT-Bench score (1–10 scale); null when LongMemEval mode was used
cases_evaluated int|null Number of evaluation cases that completed
cases_skipped_context int|null Cases skipped because context exceeded vram_cliff_tokens

Blob columns

Column Type Description
qualitative string|null Full qualitative result payload as JSON string
quantitative string|null Full quantitative result payload as JSON string
meta string|null Reproducibility hints (e.g. quality_prompts_cache_hash) as JSON string

OS / system context

Column Type Description
os_system string|null OS family: Linux, Darwin, Windows
os_release string|null Kernel / OS release string
os_machine string|null CPU architecture (e.g. x86_64, arm64)
os_distro string|null Distribution name (e.g. Ubuntu, macOS)
os_distro_version string|null Distribution version (e.g. 24.04, 15.5)
cpu_cores int|null Number of logical CPU cores
ram_total_gb float|null Total system RAM in GB

Submission metadata

Column Type Description
submitter string|null Optional public display name of the contributor
timestamp string|null ISO 8601 UTC time the benchmark run produced the row
submitted_at string|null ISO 8601 UTC time the row was uploaded

Provenance and deduplication

Column Type Description
schema_version string Schema version at time of flattening (0.9.0)
benchmark_version string PPB software version that produced the row
suite_run_id string|null UUID shared by all rows from the same ppb invocation
submission_id string|null UUID assigned during upload
row_id string UUID uniquely identifying this row
machine_fingerprint string SHA-256 of hardware profile fields (anonymous machine identity)
run_fingerprint string SHA-256 of benchmark configuration + machine fingerprint
result_fingerprint string SHA-256 of run identity + measured metrics — uniquely identifies one result
source_file_sha256 string|null SHA-256 of the source JSONL file

Extensibility

Column Type Description
tags string|null Free-form JSON string for arbitrary metadata from the suite TOML

Null value guide

Many columns are runner-specific. Expected nulls by runner type:

Column group llama-bench llama-server llama-server-loadtest
TTFT / ITL metrics null populated populated
prompt_dataset, num_prompts, n_predict null populated populated
concurrent_users null populated populated (one row per level)
gpu_pcie_gen, gpu_pcie_width null on Apple Silicon null on Apple Silicon null on Apple Silicon
unified_memory null on NVIDIA null on NVIDIA null on NVIDIA
Qualitative columns null null null

Qualitative columns are populated only when run_type is qualitative or all.


Deduplication

Use fingerprints to control for duplicates in analysis:

# Exact duplicate rows (same result, same machine, same run)
df.drop_duplicates(subset=["result_fingerprint"], inplace=True)

# Latest run per (gpu, model, quant, n_ctx, concurrent_users) config
latest = (
    df.sort_values("timestamp")
      .drop_duplicates(subset=["run_fingerprint"], keep="last")
)

Ecosystem

Benchmark tool poor-pauls-benchmark — run benchmarks and contribute results
MCP server ppb-mcp — lets any MCP-compatible LLM client query this dataset directly
Analytics poorpaul.dev/insights — leaderboard and visual analysis

Connect any MCP client to https://mcp.poorpaul.dev/mcp to query this data conversationally.


Contributing results

  1. Clone poor-pauls-benchmark
  2. Configure suites/my_gpu.toml with your hardware and models
  3. Run: uv run ppb.py all suites/my_gpu.toml
  4. Results are pushed here automatically — no PR required

No hardware contribution is too small. Every GPU tier that's missing from this dataset is a blind spot for the community.


License

Dataset content: CC BY 4.0 — contributions are attributed to their submitters.

Tooling: MIT — see the benchmark repository.

Third-party evaluation data included in rows:

Downloads last month
1,712