Support for DSpark-format speculators is being added to llama.cpp in ggml-org/llama.cpp#26275. Until merged, use a build compiled from the PR branch.

Compatibility

Component	Requirement
Target architecture	Qwen3.6 35B A3B
Recommended target	`Qwen/Qwen3.6-35B-A3B` or a compatible GGUF quantization
Draft format	DSpark
Runtime	`llama.cpp` with PR #26275 or later merged support
Standalone use	No

The target model and speculator must use compatible tokenizer, vocabulary and architecture metadata. A mismatch can cause loading errors or incorrect output.

Qwen3.6-35B-A3B Speculative Decoding Benchmark

This repository contains benchmark results for speculative decoding on Qwen3.6-35B-A3B, comparing:

No speculative decoding
DSpark with n_max = 2
DSpark with n_max = 3
DSpark with n_max = 5
DFlash with n_max = 7

The benchmark covers code generation, structured output, reasoning, factual QA, translation, creative writing, and repetitive low-entropy output.

These results are specific to the tested model, quantization, llama.cpp branch/build, hardware configuration, and benchmark prompts.
Acceptance rate alone is not sufficient to predict throughput; draft cost, accepted depth, and the number of speculative rounds also matter.

Summary

Recommended settings For Consumer GPU

Workload	Recommended setting
Code generation and completion	DFlash max 7
JSON and structured output	DFlash max 7
Repetitive or low-entropy output	DFlash max 7
Step-by-step math and constrained reasoning	DFlash max 7
General DSpark preset	DSpark max 3
Translation and creative writing	No speculative decoding
Long free-form explanation	No speculative decoding

Main findings

DFlash max 7 is the strongest overall speculative-decoding configuration.
DSpark max 5 is slower than the no-speculation baseline over the full benchmark.
Between DSpark max 2 and max 3, max 3 is the better default for code and structured tasks.
DSpark max 2 reduces some of the slowdown on explanation, translation, and long code review, but those workloads are still generally faster without speculative decoding.
Fixed long draft lengths work very well on predictable output, but waste substantial computation on high-entropy text.

Overall comparison

Aggregate wall-clock results

Mode	Predicted tokens	Draft tokens	Accepted draft tokens	Aggregate acceptance	Wall time	Effective throughput
No MTP	1890	0	0	N/A	16.69 s	113.2 tok/s
DSpark max 2	1892	1768	990	56.00%	17.43 s	108.5 tok/s
DSpark max 3	1889	2303	1102	47.85%	17.30 s	109.2 tok/s
DSpark max 5	1885	3287	1209	36.78%	18.39 s	102.5 tok/s
DFlash max 7	1885	3243	1405	43.32%	14.08 s	133.9 tok/s

Effective throughput is calculated as:

total predicted tokens / total wall-clock time

Relative to no speculative decoding

Mode	Relative performance
No MTP	1.00x
DSpark max 2	0.96x
DSpark max 3	0.96x
DSpark max 5	0.91x
DFlash max 7	1.18x

DFlash max 7 completes the benchmark approximately:

18% faster than the no-speculation baseline
31% faster than DSpark max 5

DSpark max 2 vs max 3

Per-task throughput

Task	DSpark max 2	DSpark max 3	Better
code_python	138.0	148.8	max 3
code_cpp	133.6	150.8	max 3
explain_concept	105.3	90.0	max 2
summarize	114.8	120.9	max 3
qa_factual	122.4	126.3	max 3
translation	89.0	82.6	max 2
creative_short	85.3	90.5	max 3
stepwise_math	129.5	133.0	max 3
json_output	155.3	164.3	max 3
long_reasoning	112.6	111.8	effectively tied
repeat_pattern	168.2	195.6	max 3
code_completion	140.6	154.5	max 3
long_code_review	101.8	90.7	max 2

Aggregate comparison

Metric	max 2	max 3
Total predicted tokens	1892	1889
Total draft tokens	1768	2303
Accepted draft tokens	990	1102
Aggregate acceptance	56.00%	47.85%
Wall time	17.43 s	17.30 s
Effective throughput	108.5 tok/s	109.2 tok/s

The aggregate difference is very small:

max 3 is only about 0.6% faster overall
max 3 produces 535 more draft tokens
max 3 accepts 112 more draft tokens

This means the third draft position is close to the overall break-even point.

Recommendation

Use:

--spec-draft-n-max 3

as the default DSpark setting when the workload is mainly:

source code
code completion
JSON
structured output
repetitive output
constrained reasoning

Use max 2 only when reducing the penalty on less predictable natural-language output is more important.

However, translation, creative writing, and long free-form explanations are still usually better with speculative decoding disabled.

DSpark max 5 vs DFlash max 7

Per-task throughput

Task	No MTP	DSpark max 5	DFlash max 7
code_python	128.8	133.0	206.1
code_cpp	129.6	172.5	229.5
explain_concept	129.6	82.5	98.5
summarize	128.6	113.9	147.9
qa_factual	129.1	112.2	146.3
translation	126.3	71.1	68.8
creative_short	128.7	70.5	73.7
stepwise_math	130.9	131.4	184.3
json_output	129.5	175.7	237.4
long_reasoning	127.4	113.2	173.2
repeat_pattern	129.3	239.9	300.9
code_completion	125.3	178.8	275.9
long_code_review	126.5	82.3	96.8

Acceptance comparison

Task	DSpark max 5	DFlash max 7
code_python	45.1%	65.3%
code_cpp	64.6%	71.4%
explain_concept	17.9%	19.9%
summarize	32.6%	36.7%
qa_factual	30.7%	36.9%
translation	14.0%	11.4%
creative_short	10.9%	11.4%
stepwise_math	41.2%	53.8%
json_output	64.2%	73.7%
long_reasoning	31.4%	45.0%
repeat_pattern	92.3%	97.1%
code_completion	65.2%	94.3%
long_code_review	17.8%	20.3%

DFlash generates slightly fewer total draft tokens than DSpark max 5, while accepting substantially more:

Metric	DSpark max 5	DFlash max 7
Total draft tokens	3287	3243
Accepted draft tokens	1209	1405
Aggregate acceptance	36.78%	43.32%
Wall time	18.39 s	14.08 s

This suggests that DFlash is not winning by brute-force drafting. Its proposals are better aligned with the target model, resulting in:

deeper accepted runs
fewer speculative rounds
fewer wasted proposals
better utilization of longer draft sequences

Workload analysis

Code and structured output

DFlash max 7 is clearly the strongest configuration.

Task	DFlash speedup vs baseline
code_python	1.60x
code_cpp	1.77x
json_output	1.83x
code_completion	2.20x
repeat_pattern	2.33x
stepwise_math	1.41x

These tasks are predictable enough for long speculative runs to be accepted frequently.

In particular:

code_completion reaches 94.3% acceptance
repeat_pattern reaches 97.1% acceptance
json_output reaches 73.7% acceptance

This is the ideal operating region for speculative decoding.

Factual QA and reasoning

DFlash remains useful on several medium-entropy tasks:

Task	Baseline	DFlash max 7	Speedup
summarize	128.6	147.9	1.15x
qa_factual	129.1	146.3	1.13x
long_reasoning	127.4	173.2	1.36x

DSpark max 5 does not reach the same break-even point on these workloads.

Translation and creative writing

Both DSpark and DFlash perform poorly here.

Task	Baseline	DSpark max 5	DFlash max 7
translation	126.3	71.1	68.8
creative_short	128.7	70.5	73.7

Acceptance falls to approximately 11-14%, so most later draft positions are wasted.

For these workloads, use:

No speculative decoding

Explanation and long code review

Long explanatory text and code-review prose also remain difficult:

Task	Baseline	DFlash max 7
explain_concept	129.6	98.5
long_code_review	126.5	96.8

Although the subject may be technical, the output itself is less constrained than code completion. The next token distribution remains too broad for a fixed seven-token draft chain to be consistently useful.

Why aggregate acceptance rate is not enough

A single aggregate acceptance percentage hides positional behavior.

For example, two configurations may have similar aggregate acceptance while behaving very differently:

one may frequently accept three or four tokens per round
another may accept only the first token in many rounds
one may perform fewer total speculative rounds
another may repeatedly compute later draft positions that are rejected

The most useful additional metric would be positional acceptance:

P1, P2, P3, P4, P5, P6, P7

This would expose the acceptance-decay knee directly.

Expected behavior from these results:

DFlash likely maintains stronger acceptance at later positions on code and JSON.
DSpark likely decays earlier, especially after the second or third position.
Translation and creative writing likely collapse immediately after the first position.

Dynamic confidence truncation

A fixed n_max applies the same draft depth to every speculative round, regardless of task entropy.

A better implementation would stop draft generation dynamically when confidence becomes too low:

text

if confidence[position] < threshold:
    stop drafting

This could allow:

code completion to use all seven positions
JSON to use five to seven positions
factual QA to stop after two to four positions
translation and creative output to stop after one position
very uncertain rounds to skip speculative decoding entirely

Dynamic early exit could preserve the large DFlash gains on structured workloads while reducing severe regressions on free-form text.

Raw benchmark results

DSpark max 2

text

code_python        pred= 192 draft= 160 acc= 110 rate=0.688 tok/s=138.0
code_cpp           pred=  60 draft=  52 acc=  35 rate=0.673 tok/s=133.6
explain_concept    pred= 192 draft= 220 acc=  80 rate=0.364 tok/s=105.3
summarize          pred=  51 draft=  52 acc=  24 rate=0.462 tok/s=114.8
qa_factual         pred= 192 draft= 188 acc=  96 rate=0.511 tok/s=122.4
translation        pred=  17 draft=  20 acc=   6 rate=0.300 tok/s=89.0
creative_short     pred=  36 draft=  50 acc=  10 rate=0.200 tok/s=85.3
stepwise_math      pred= 192 draft= 173 acc= 103 rate=0.595 tok/s=129.5
json_output        pred= 192 draft= 141 acc= 120 rate=0.851 tok/s=155.3
long_reasoning     pred= 192 draft= 204 acc=  89 rate=0.436 tok/s=112.6
repeat_pattern     pred= 192 draft= 131 acc= 125 rate=0.954 tok/s=168.2
code_completion    pred= 192 draft= 157 acc= 112 rate=0.713 tok/s=140.6
long_code_review   pred= 192 draft= 220 acc=  80 rate=0.364 tok/s=101.8

json

{
  "n_requests": 13,
  "total_predicted": 1892,
  "total_draft": 1768,
  "total_draft_accepted": 990,
  "aggregate_accept_rate": 0.56,
  "wall_s_total": 17.43
}

DSpark max 3

text

code_python        pred= 192 draft= 195 acc= 126 rate=0.646 tok/s=148.8
code_cpp           pred=  54 draft=  54 acc=  37 rate=0.685 tok/s=150.8
explain_concept    pred= 192 draft= 330 acc=  79 rate=0.239 tok/s=90.0
summarize          pred=  52 draft=  63 acc=  30 rate=0.476 tok/s=120.9
qa_factual         pred= 192 draft= 240 acc= 110 rate=0.458 tok/s=126.3
translation        pred=  17 draft=  30 acc=   6 rate=0.200 tok/s=82.6
creative_short     pred=  38 draft=  66 acc=  15 rate=0.227 tok/s=90.5
stepwise_math      pred= 192 draft= 222 acc= 116 rate=0.522 tok/s=133.0
json_output        pred= 192 draft= 175 acc= 132 rate=0.754 tok/s=164.3
long_reasoning     pred= 192 draft= 264 acc= 102 rate=0.386 tok/s=111.8
repeat_pattern     pred= 192 draft= 147 acc= 141 rate=0.959 tok/s=195.6
code_completion    pred= 192 draft= 187 acc= 128 rate=0.684 tok/s=154.5
long_code_review   pred= 192 draft= 330 acc=  80 rate=0.242 tok/s=90.7

json

{
  "n_requests": 13,
  "total_predicted": 1889,
  "total_draft": 2303,
  "total_draft_accepted": 1102,
  "aggregate_accept_rate": 0.4785,
  "wall_s_total": 17.3
}

DSpark max 5

text

code_python        pred= 192 draft= 293 acc= 132 rate=0.451 tok/s=133.0
code_cpp           pred=  54 draft=  65 acc=  42 rate=0.646 tok/s=172.5
explain_concept    pred= 192 draft= 498 acc=  89 rate=0.179 tok/s=82.5
summarize          pred=  51 draft=  95 acc=  31 rate=0.326 tok/s=113.9
qa_factual         pred= 192 draft= 375 acc= 115 rate=0.307 tok/s=112.2
translation        pred=  17 draft=  50 acc=   7 rate=0.140 tok/s=71.1
creative_short     pred=  35 draft= 110 acc=  12 rate=0.109 tok/s=70.5
stepwise_math      pred= 192 draft= 311 acc= 128 rate=0.412 tok/s=131.4
json_output        pred= 192 draft= 226 acc= 145 rate=0.642 tok/s=175.7
long_reasoning     pred= 192 draft= 370 acc= 116 rate=0.314 tok/s=113.2
repeat_pattern     pred= 192 draft= 170 acc= 157 rate=0.923 tok/s=239.9
code_completion    pred= 192 draft= 224 acc= 146 rate=0.652 tok/s=178.8
long_code_review   pred= 192 draft= 500 acc=  89 rate=0.178 tok/s=82.3

json

{
  "n_requests": 13,
  "total_predicted": 1885,
  "total_draft": 3287,
  "total_draft_accepted": 1209,
  "aggregate_accept_rate": 0.3678,
  "wall_s_total": 18.39
}

DFlash max 7

text

code_python        pred= 192 draft= 239 acc= 156 rate=0.653 tok/s=206.1
code_cpp           pred=  54 draft=  63 acc=  45 rate=0.714 tok/s=229.5
explain_concept    pred= 192 draft= 552 acc= 110 rate=0.199 tok/s=98.5
summarize          pred=  51 draft=  98 acc=  36 rate=0.367 tok/s=147.9
qa_factual         pred= 192 draft= 371 acc= 137 rate=0.369 tok/s=146.3
translation        pred=  17 draft=  70 acc=   8 rate=0.114 tok/s=68.8
creative_short     pred=  35 draft= 140 acc=  16 rate=0.114 tok/s=73.7
stepwise_math      pred= 192 draft= 279 acc= 150 rate=0.538 tok/s=184.3
json_output        pred= 192 draft= 217 acc= 160 rate=0.737 tok/s=237.4
long_reasoning     pred= 192 draft= 322 acc= 145 rate=0.450 tok/s=173.2
repeat_pattern     pred= 192 draft= 171 acc= 166 rate=0.971 tok/s=300.9
code_completion    pred= 192 draft= 175 acc= 165 rate=0.943 tok/s=275.9
long_code_review   pred= 192 draft= 546 acc= 111 rate=0.203 tok/s=96.8

json

{
  "n_requests": 13,
  "total_predicted": 1885,
  "total_draft": 3243,
  "total_draft_accepted": 1405,
  "aggregate_accept_rate": 0.4332,
  "wall_s_total": 14.08
}

No speculative decoding

text

code_python        pred= 192 draft=   0 acc=   0 rate=n/a tok/s=128.8
code_cpp           pred=  54 draft=   0 acc=   0 rate=n/a tok/s=129.6
explain_concept    pred= 192 draft=   0 acc=   0 rate=n/a tok/s=129.6
summarize          pred=  51 draft=   0 acc=   0 rate=n/a tok/s=128.6
qa_factual         pred= 192 draft=   0 acc=   0 rate=n/a tok/s=129.1
translation        pred=  17 draft=   0 acc=   0 rate=n/a tok/s=126.3
creative_short     pred=  40 draft=   0 acc=   0 rate=n/a tok/s=128.7
stepwise_math      pred= 192 draft=   0 acc=   0 rate=n/a tok/s=130.9
json_output        pred= 192 draft=   0 acc=   0 rate=n/a tok/s=129.5
long_reasoning     pred= 192 draft=   0 acc=   0 rate=n/a tok/s=127.4
repeat_pattern     pred= 192 draft=   0 acc=   0 rate=n/a tok/s=129.3
code_completion    pred= 192 draft=   0 acc=   0 rate=n/a tok/s=125.3
long_code_review   pred= 192 draft=   0 acc=   0 rate=n/a tok/s=126.5

json

{
  "n_requests": 13,
  "total_predicted": 1890,
  "total_draft": 0,
  "total_draft_accepted": 0,
  "aggregate_accept_rate": null,
  "wall_s_total": 16.69
}

Final recommendation

For a single general DSpark preset:

--spec-draft-n-max 3

For maximum performance on code, JSON, agent output, and other predictable workloads:

DFlash with --spec-draft-n-max 7

For translation, creative writing, roleplay, and long free-form explanations:

Disable speculative decoding

The benchmark strongly suggests that the ideal long-term solution is not a single fixed draft length, but confidence-based dynamic draft truncation.

williamliao/Qwen3.6-35B-A3B-DSPARK-GGUF

Qwen3.6-35B-A3B-DSPARK-GGUF

Model

Status

Compatibility

Qwen3.6-35B-A3B Speculative Decoding Benchmark

Summary

Recommended settings For Consumer GPU

Main findings

Overall comparison

Aggregate wall-clock results

Relative to no speculative decoding

DSpark max 2 vs max 3

Per-task throughput

Aggregate comparison

Recommendation

DSpark max 5 vs DFlash max 7

Per-task throughput

Acceptance comparison

Workload analysis

Code and structured output

Factual QA and reasoning

Translation and creative writing

Explanation and long code review

Why aggregate acceptance rate is not enough

Dynamic confidence truncation

Raw benchmark results

DSpark max 2

DSpark max 3

DSpark max 5

DFlash max 7

No speculative decoding

Final recommendation

williamliao/
Qwen3.6-35B-A3B-DSPARK-GGUF