Benchmarking Qwen3.6-27B MTP Performance
RTX 5090, Ryzen 9 9950X3D, 128GB DDR5
Debian 13 6.12.85-1, CUDA 13.2.78, llama.cpp b9200
1. Control - No MTP bits in gguf
model
test
t/s
peak t/s
ttfr (ms)
est_ppt (ms)
e2e_ttft (ms)
unsloth/Qwen3.6-27B
pp2048
2642.44 ± 58.11
782.48 ± 17.10
775.67 ± 17.10
782.48 ± 17.10
unsloth/Qwen3.6-27B
tg32
56.75 ± 2.09
58.66 ± 2.16
model
test
t/s
peak t/s
ttfr (ms)
est_ppt (ms)
e2e_ttft (ms)
unsloth/Qwen3.6-27B
pp2048
2742.91 ± 206.98
758.36 ± 55.83
751.22 ± 55.83
758.36 ± 55.83
unsloth/Qwen3.6-27B
tg32
55.48 ± 3.51
57.34 ± 3.63
2. MTP bits present but not enabled
model
test
t/s
peak t/s
ttfr (ms)
est_ppt (ms)
e2e_ttft (ms)
unsloth/Qwen3.6-27B-Q4
pp2048
3007.29 ± 192.09
618.65 ± 62.49
611.74 ± 62.49
618.65 ± 62.49
unsloth/Qwen3.6-27B-Q4
tg32
83.80 ± 7.19
86.40 ± 7.38
unsloth/Qwen3.6-27B-Q5
pp2048
2742.71 ± 161.67
695.08 ± 49.64
686.56 ± 49.64
695.08 ± 49.64
unsloth/Qwen3.6-27B-Q5
tg32
85.52 ± 5.58
88.14 ± 5.83
unsloth/Qwen3.6-27B-Q6
pp2048
2464.15 ± 133.45
732.50 ± 55.54
724.57 ± 55.54
732.50 ± 55.54
unsloth/Qwen3.6-27B-Q6
tg32
66.02 ± 3.63
68.02 ± 3.74
model
test
t/s
peak t/s
ttfr (ms)
est_ppt (ms)
e2e_ttft (ms)
unsloth/Qwen3.6-27B-Q4
pp2048
2945.08 ± 292.27
638.99 ± 59.82
630.82 ± 59.82
638.99 ± 59.82
unsloth/Qwen3.6-27B-Q4
tg32
84.88 ± 7.24
87.49 ± 7.43
unsloth/Qwen3.6-27B-Q5
pp2048
2701.85 ± 176.83
687.89 ± 44.52
681.62 ± 44.52
687.89 ± 44.52
unsloth/Qwen3.6-27B-Q5
tg32
85.14 ± 11.00
87.79 ± 11.36
unsloth/Qwen3.6-27B-Q6
pp2048
2548.04 ± 246.37
752.56 ± 50.19
746.01 ± 50.19
752.56 ± 50.19
unsloth/Qwen3.6-27B-Q6
tg32
70.57 ± 5.47
72.58 ± 5.52
3. MTP bits enabled in llama.cpp config
model
test
t/s
peak t/s
ttfr (ms)
est_ppt (ms)
e2e_ttft (ms)
unsloth/Qwen3.6-27B-Q4-MTP
pp2048
2740.98 ± 91.47
697.70 ± 15.55
689.78 ± 15.55
697.70 ± 15.55
unsloth/Qwen3.6-27B-Q4-MTP
tg32
186.58 ± 39.52
193.47 ± 41.12
unsloth/Qwen3.6-27B-Q5-MTP
pp2048
340.37 ± 9.89
5545.89 ± 292.50
5538.65 ± 292.50
5545.89 ± 292.50
unsloth/Qwen3.6-27B-Q5-MTP
tg32
80.81 ± 10.34
83.35 ± 10.67
unsloth/Qwen3.6-27B-Q6-MTP
pp2048
206.50 ± 1.71
9086.85 ± 411.00
9079.60 ± 411.00
9086.85 ± 411.00
unsloth/Qwen3.6-27B-Q6-MTP
tg32
36.32 ± 1.19
37.42 ± 1.23
model
test
t/s
peak t/s
ttfr (ms)
est_ppt (ms)
e2e_ttft (ms)
unsloth/Qwen3.6-27B-Q4-MTP
pp2048
2623.13 ± 109.43
709.12 ± 29.69
701.49 ± 29.69
709.12 ± 29.69
unsloth/Qwen3.6-27B-Q4-MTP
tg32
238.85 ± 30.30
247.89 ± 31.85
unsloth/Qwen3.6-27B-Q5-MTP
pp2048
357.79 ± 12.51
5185.61 ± 40.74
5176.51 ± 40.74
5185.61 ± 40.74
unsloth/Qwen3.6-27B-Q5-MTP
tg32
93.51 ± 19.10
96.41 ± 19.85
unsloth/Qwen3.6-27B-Q6-MTP
pp2048
211.54 ± 8.20
8909.21 ± 525.73
8902.88 ± 525.73
8909.21 ± 525.73
unsloth/Qwen3.6-27B-Q6-MTP
tg32
42.67 ± 8.76
43.97 ± 9.05
All tests run with MTP and mmproj enabled.
unsloth/Qwen3.6-27B-Q4-MTP
Task
Predicted
Draft
Accepted
Accept Rate
Tokens/s
code_python
192
168
134
0.798
138.8
code_cpp
192
195
125
0.641
120.9
explain_concept
192
188
127
0.675
124.2
summarize
192
162
136
0.840
143.1
qa_factual
192
181
130
0.718
129.3
translation
192
209
121
0.579
114.1
creative_short
192
214
119
0.556
111.0
stepwise_math
192
173
133
0.769
135.6
long_code_review
192
209
121
0.579
112.6
Aggregate: 9 requests | 1728 predicted | 1699 draft | 1146 accepted | Accept rate: 67.45% | Wall time: 15.59s
unsloth/Qwen3.6-27B-Q5-MTP
Task
Predicted
Draft
Accepted
Accept Rate
Tokens/s
code_python
192
166
135
0.813
55.8
code_cpp
192
201
124
0.617
46.8
explain_concept
192
192
126
0.656
48.6
summarize
192
162
136
0.840
56.8
qa_factual
192
198
125
0.631
47.7
translation
192
196
125
0.638
47.6
creative_short
192
199
124
0.623
47.0
stepwise_math
192
177
132
0.746
53.3
long_code_review
192
204
121
0.593
44.8
Aggregate: 9 requests | 1728 predicted | 1695 draft | 1148 accepted | Accept rate: 67.73% | Wall time: 41.33s
unsloth/Qwen3.6-27B-Q6-MTP
Task
Predicted
Draft
Accepted
Accept Rate
Tokens/s
code_python
192
159
137
0.862
34.9
code_cpp
192
189
126
0.667
28.5
explain_concept
192
197
125
0.634
28.6
summarize
192
168
134
0.798
33.2
qa_factual
192
172
132
0.767
32.1
translation
192
192
126
0.656
29.2
creative_short
192
207
122
0.589
27.3
stepwise_math
192
176
132
0.750
31.7
long_code_review
192
202
123
0.609
27.4
Aggregate: 9 requests | 1728 predicted | 1662 draft | 1157 accepted | Accept rate: 69.61% | Wall time: 67.29s