Views
No views yet
This is a DFlash draft model, not a standalone language model.It must be used together with a compatible google/gemma-4-12B-it GGUF target model.
z-lab/gemma4-12B-it-DFlashgoogle/gemma-4-12B-itgemma-4-12b-it-UD-Q4_K_XL1llama-server \
2 -m gemma-4-12b-it-UD-Q4_K_XL.gguf \
3 -md gemma4-12B-it-DFlash-Q4_K_M.gguf \
4 --spec-type draft-dflash \
5 --spec-draft-n-max 41llama-cli \
2 -m gemma-4-12b-it-UD-Q4_K_XL.gguf \
3 -md gemma4-12B-it-DFlash-Q4_K_M.gguf \
4 --spec-type draft-dflash \
5 --spec-draft-n-max 4n_max = 4 gave the best aggregate wall-clock result.n_max = 2 gave the highest overall acceptance rate and performed better on several high-entropy natural-language tasks.n_max = 4–5 performed well for predictable outputs such as code completion, JSON, and repeated patterns.n_max = 5 improved a few highly structured tasks, but was slower overall than n_max = 4 in this mixed benchmark.--spec-draft-n-max 4n_max = 2 may be preferable. For JSON and highly predictable output, n_max = 4–5 may perform better. For general code completion, n_max = 3–4 may be a better starting point.convert_hf_to_gguf.py.--target-model-dir.gemma-4-12b-it-UD-Q4_K_XLgemma4-12B-it-DFlash-Q4_K_Mn_max | Draft acceptance | Predicted tokens | Draft tokens | Accepted tokens | Wall time |
|---|---|---|---|---|---|
| 2 | 67.6% | 1,975 | 1,668 | 1,128 | 19.50 s |
| 3 | 57.5% | 1,975 | 2,159 | 1,241 | 19.29 s |
| 4 | 51.0% | 1,976 | 2,582 | 1,316 | 18.81 s |
| 5 | 42.7% | 1,976 | 3,125 | 1,334 | 20.45 s |
n_max = 21code_python pred= 192 draft= 157 acc= 112 rate=0.713 tok/s=112.2
2code_cpp pred= 192 draft= 152 acc= 115 rate=0.757 tok/s=123.0
3explain_concept pred= 192 draft= 190 acc= 95 rate=0.500 tok/s=98.1
4summarize pred= 43 draft= 38 acc= 24 rate=0.632 tok/s=110.2
5qa_factual pred= 160 draft= 148 acc= 85 rate=0.574 tok/s=104.5
6translation pred= 26 draft= 32 acc= 10 rate=0.312 tok/s=78.2
7creative_short pred= 44 draft= 56 acc= 15 rate=0.268 tok/s=77.0
8stepwise_math pred= 192 draft= 145 acc= 118 rate=0.814 tok/s=128.3
9json_output pred= 192 draft= 134 acc= 124 rate=0.925 tok/s=134.3
10long_reasoning pred= 192 draft= 155 acc= 113 rate=0.729 tok/s=115.2
11repeat_pattern pred= 192 draft= 130 acc= 126 rate=0.969 tok/s=141.6
12code_completion pred= 166 draft= 128 acc= 102 rate=0.797 tok/s=125.2
13long_code_review pred= 192 draft= 203 acc= 89 rate=0.438 tok/s=88.7
14
15Aggregate:
16 requests: 13
17 predicted tokens: 1975
18 draft tokens: 1668
19 accepted tokens: 1128
20 acceptance rate: 0.6763
21 total wall time: 19.50 sn_max = 31code_python pred= 192 draft= 206 acc= 122 rate=0.592 tok/s=114.0
2code_cpp pred= 192 draft= 197 acc= 125 rate=0.634 tok/s=122.0
3explain_concept pred= 192 draft= 261 acc= 103 rate=0.395 tok/s=92.3
4summarize pred= 43 draft= 48 acc= 27 rate=0.562 tok/s=112.3
5qa_factual pred= 160 draft= 201 acc= 93 rate=0.463 tok/s=98.0
6translation pred= 26 draft= 45 acc= 10 rate=0.222 tok/s=69.8
7creative_short pred= 44 draft= 78 acc= 18 rate=0.231 tok/s=70.4
8stepwise_math pred= 192 draft= 186 acc= 129 rate=0.694 tok/s=130.2
9json_output pred= 192 draft= 155 acc= 139 rate=0.897 tok/s=152.4
10long_reasoning pred= 192 draft= 195 acc= 125 rate=0.641 tok/s=122.0
11repeat_pattern pred= 192 draft= 146 acc= 142 rate=0.973 tok/s=162.2
12code_completion pred= 166 draft= 156 acc= 114 rate=0.731 tok/s=132.7
13long_code_review pred= 192 draft= 285 acc= 94 rate=0.330 tok/s=79.3
14
15Aggregate:
16 requests: 13
17 predicted tokens: 1975
18 draft tokens: 2159
19 accepted tokens: 1241
20 acceptance rate: 0.5748
21 total wall time: 19.29 sn_max = 41code_python pred= 192 draft= 257 acc= 126 rate=0.490 tok/s=112.7
2code_cpp pred= 192 draft= 220 acc= 135 rate=0.614 tok/s=132.7
3explain_concept pred= 192 draft= 304 acc= 114 rate=0.375 tok/s=96.4
4summarize pred= 46 draft= 68 acc= 29 rate=0.426 tok/s=102.9
5qa_factual pred= 162 draft= 248 acc= 100 rate=0.403 tok/s=101.8
6translation pred= 26 draft= 56 acc= 12 rate=0.214 tok/s=70.8
7creative_short pred= 44 draft= 100 acc= 19 rate=0.190 tok/s=68.8
8stepwise_math pred= 192 draft= 219 acc= 136 rate=0.621 tok/s=133.6
9json_output pred= 192 draft= 176 acc= 146 rate=0.830 tok/s=163.8
10long_reasoning pred= 192 draft= 228 acc= 133 rate=0.583 tok/s=127.3
11repeat_pattern pred= 192 draft= 163 acc= 149 rate=0.914 tok/s=174.7
12code_completion pred= 162 draft= 196 acc= 113 rate=0.577 tok/s=127.3
13long_code_review pred= 192 draft= 347 acc= 104 rate=0.300 tok/s=80.2
14
15Aggregate:
16 requests: 13
17 predicted tokens: 1976
18 draft tokens: 2582
19 accepted tokens: 1316
20 acceptance rate: 0.5097
21 total wall time: 18.81 sn_max = 51code_python pred= 192 draft= 285 acc= 133 rate=0.467 tok/s=112.0
2code_cpp pred= 192 draft= 285 acc= 134 rate=0.470 tok/s=116.8
3explain_concept pred= 192 draft= 384 acc= 113 rate=0.294 tok/s=85.5
4summarize pred= 46 draft= 85 acc= 29 rate=0.341 tok/s=92.6
5qa_factual pred= 162 draft= 310 acc= 100 rate=0.323 tok/s=90.3
6translation pred= 26 draft= 70 acc= 12 rate=0.171 tok/s=62.9
7creative_short pred= 44 draft= 125 acc= 19 rate=0.152 tok/s=58.6
8stepwise_math pred= 192 draft= 257 acc= 139 rate=0.541 tok/s=124.4
9json_output pred= 192 draft= 182 acc= 154 rate=0.846 tok/s=171.4
10long_reasoning pred= 192 draft= 274 acc= 135 rate=0.493 tok/s=115.3
11repeat_pattern pred= 192 draft= 173 acc= 155 rate=0.896 tok/s=175.2
12code_completion pred= 162 draft= 225 acc= 117 rate=0.520 tok/s=120.2
13long_code_review pred= 192 draft= 470 acc= 94 rate=0.200 tok/s=67.8
14
15Aggregate:
16 requests: 13
17 predicted tokens: 1976
18 draft tokens: 3125
19 accepted tokens: 1334
20 acceptance rate: 0.4269
21 total wall time: 20.45 sn_max = 2 achieved the highest aggregate draft acceptance rate.n_max = 4 achieved the shortest total wall time in this mixed benchmark.n_max = 3 was close to n_max = 4 while using fewer draft tokens.n_max = 5 improved JSON output and repeated-pattern generation, but increased total wall time.n_max increased.n_max = 3 in this benchmark, while JSON and repeated patterns benefited from longer draft sequences.google/gemma-4-12B-it GGUF target model is required. The target GGUF may use a different quantization from the draft model, but both models must share compatible tokenizer and model architecture assumptions.google/gemma-4-12B-itz-lab/gemma4-12B-it-DFlash and google/gemma-4-12B-it repositories identify their applicable licensing terms. Users should review the upstream model cards and licenses before redistribution or commercial use.