Qwen3.6 35B-A3B
Qwen/Qwen3.6-35B-A3B
Qwen · 9 recorded configurations
Measured setup
Hardware
Runtime and tool
Quantisation
MTP draft depth
Every option opens a recorded setup. Where the corpus recorded nothing at this page's other settings, the option names the fields it switches with it. Allocation and runtime detail stay in the technical record below.
1× RTX 3090 · llama.cpp
UD-Q3_K_XL · MTP 2
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 2,610.33 | 226.53 |
| 2,048 | 2,169.61 | 221.71 |
| 4,096 | 2,087.39 | 218.29 |
| 8,192 | 1,919.55 | 211.64 |
| 16,384 | 1,681.95 | 202.44 |
| 32,768 | 1,345.36 | 186.39 |
| 65,536 | 1,000.92 | 160.17 |
| 131,072 | 641.74 | 124.55 |
| 261,312 | 369.43 | 87.02 |
Concurrent-stream measurements
| Streams | Context depth (tokens) | Aggregate PP (tok/s) | Aggregate TG (tok/s) | TG / stream (tok/s) | Aggregate all (tok/s) |
|---|---|---|---|---|---|
| 2 | 8,192 | 1,836.66 | 275.19 | 137.59 | 645.37 |
| 3 | 8,192 | 1,688.19 | 278.72 | 92.91 | 694.90 |
| 4 | 8,192 | 1,507.03 | 290.72 | 72.68 | 745.83 |
| 5 | 8,192 | 1,522.64 | 309.33 | 61.87 | 810.11 |
| 6 | 8,192 | 1,497.68 | 314.83 | 52.47 | 841.21 |
| 7 | 8,192 | 1,489.25 | 323.35 | 46.19 | 875.02 |
| 8 | 8,192 | 1,499.16 | 332.10 | 41.51 | 909.93 |
Technical record and provenance
Source 1 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-context-2026-10-14-single-gpu-isolated-mtp-2-c1
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-2-f3ad681bf9
- Entry sweep
- context
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 262,144 tokens
- Actual context per slot
- 262,144 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 2
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 0, 1 stream: PP mean 2,610.33 t/s (min 2,608.13, max 2,613.04, SD 2.04, n=3); TG mean 226.53 t/s (min 226.33, max 226.90, SD 0.26, n=3); All mean 565.37 t/s (min 564.78, max 566.44, SD 0.76, n=3) · KV 2,048, 1 stream: PP mean 2,169.61 t/s (min 2,166.28, max 2,176.04, SD 4.55, n=3); TG mean 221.71 t/s (min 221.32, max 221.93, SD 0.28, n=3); All mean 539.36 t/s (min 538.78, max 539.72, SD 0.41, n=3) · KV 4,096, 1 stream: PP mean 2,087.39 t/s (min 2,035.07, max 2,121.86, SD 37.61, n=3); TG mean 218.29 t/s (min 218.01, max 218.70, SD 0.29, n=3); All mean 528.89 t/s (min 526.07, max 531.08, SD 2.10, n=3) · KV 8,192, 1 stream: PP mean 1,919.55 t/s (min 1,887.99, max 1,943.06, SD 23.19, n=3); TG mean 211.64 t/s (min 210.67, max 212.37, SD 0.72, n=3); All mean 508.12 t/s (min 506.99, max 509.88, SD 1.26, n=3) · KV 16,384, 1 stream: PP mean 1,681.95 t/s (min 1,638.47, max 1,713.11, SD 31.69, n=3); TG mean 202.44 t/s (min 200.16, max 203.66, SD 1.61, n=3); All mean 477.91 t/s (min 471.44, max 481.70, SD 4.60, n=3) · KV 32,768, 1 stream: PP mean 1,345.36 t/s (min 1,293.35, max 1,379.32, SD 37.34, n=3); TG mean 186.39 t/s (min 185.52, max 187.06, SD 0.64, n=3); All mean 427.13 t/s (min 421.84, max 429.82, SD 3.74, n=3) · KV 65,536, 1 stream: PP mean 1,000.92 t/s (min 997.44, max 1,003.13, SD 2.49, n=3); TG mean 160.17 t/s (min 158.45, max 161.10, SD 1.22, n=3); All mean 355.22 t/s (min 352.06, max 357.03, SD 2.25, n=3) · KV 131,072, 1 stream: PP mean 641.74 t/s (min 639.67, max 643.80, SD 1.68, n=3); TG mean 124.55 t/s (min 123.91, max 125.44, SD 0.65, n=3); All mean 262.38 t/s (min 261.21, max 264.18, SD 1.29, n=3) · KV 261,312, 1 stream: PP mean 369.43 t/s (min 357.19, max 376.09, SD 8.67, n=3); TG mean 87.02 t/s (min 86.84, max 87.18, SD 0.14, n=3); All mean 173.04 t/s (min 171.16, max 174.00, SD 1.34, n=3)
- Measurement outcomes
- 0 tokens · 1 stream: measured; 2,048 tokens · 1 stream: measured; 4,096 tokens · 1 stream: measured; 8,192 tokens · 1 stream: measured; 16,384 tokens · 1 stream: measured; 32,768 tokens · 1 stream: measured; 65,536 tokens · 1 stream: measured; 131,072 tokens · 1 stream: measured; 261,312 tokens · 1 stream: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-2, concurrency=1). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-14-approved-models-single-gpu/qwen36-q3-k-xl-context-isolated
Source 2 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-mtp-2-c2
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-2-f3ad681bf9
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 2
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 2 streams: PP mean 1,836.66 t/s (min 1,740.35, max 1,985.96, SD 107.04, n=3); TG mean 275.19 t/s (min 273.51, max 277.11, SD 1.48, n=3); All mean 645.37 t/s (min 635.45, max 659.62, SD 10.33, n=3)
- Measurement outcomes
- 8,192 tokens · 2 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-2, concurrency=2). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 3 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-mtp-2-c3
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-2-f3ad681bf9
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 2
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 3 streams: PP mean 1,688.19 t/s (min 1,610.17, max 1,743.76, SD 56.81, n=3); TG mean 278.72 t/s (min 275.55, max 282.11, SD 2.68, n=3); All mean 694.90 t/s (min 683.21, max 704.23, SD 8.74, n=3)
- Measurement outcomes
- 8,192 tokens · 3 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-2, concurrency=3). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 4 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-mtp-2-c4
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-2-f3ad681bf9
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 2
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 4 streams: PP mean 1,507.03 t/s (min 1,485.71, max 1,533.47, SD 19.83, n=3); TG mean 290.72 t/s (min 286.74, max 296.73, SD 4.32, n=3); All mean 745.83 t/s (min 733.95, max 758.98, SD 10.26, n=3)
- Measurement outcomes
- 8,192 tokens · 4 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-2, concurrency=4). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 5 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-mtp-2-c5
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-2-f3ad681bf9
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 2
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 5 streams: PP mean 1,522.64 t/s (min 1,505.44, max 1,553.93, SD 22.16, n=3); TG mean 309.33 t/s (min 306.77, max 311.90, SD 2.10, n=3); All mean 810.11 t/s (min 801.71, max 818.56, SD 6.88, n=3)
- Measurement outcomes
- 8,192 tokens · 5 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-2, concurrency=5). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 6 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-mtp-2-c6
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-2-f3ad681bf9
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 2
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 6 streams: PP mean 1,497.68 t/s (min 1,478.32, max 1,508.01, SD 13.70, n=3); TG mean 314.83 t/s (min 313.50, max 316.52, SD 1.26, n=3); All mean 841.21 t/s (min 839.01, max 845.48, SD 3.02, n=3)
- Measurement outcomes
- 8,192 tokens · 6 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-2, concurrency=6). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 7 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-mtp-2-c7
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-2-f3ad681bf9
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 2
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 7 streams: PP mean 1,489.25 t/s (min 1,465.81, max 1,522.29, SD 24.03, n=3); TG mean 323.35 t/s (min 320.99, max 326.03, SD 2.07, n=3); All mean 875.02 t/s (min 867.53, max 884.01, SD 6.81, n=3)
- Measurement outcomes
- 8,192 tokens · 7 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-2, concurrency=7). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0
Source 8 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-q3-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu0-co-scheduled-mtp-2-c8
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-2-f3ad681bf9
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-Q3_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 262,144 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 2
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 8 streams: PP mean 1,499.16 t/s (min 1,440.48, max 1,534.76, SD 41.81, n=3); TG mean 332.10 t/s (min 329.22, max 334.36, SD 2.14, n=3); All mean 909.93 t/s (min 894.87, max 918.24, SD 10.66, n=3)
- Measurement outcomes
- 8,192 tokens · 8 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-2, concurrency=8). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/remaining-campaign/qwen36-q3-k-xl-concurrency-physical0