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 3
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 2,619.87 | 243.73 |
| 2,048 | 2,134.34 | 238.32 |
| 4,096 | 2,132.94 | 237.02 |
| 8,192 | 2,003.78 | 231.90 |
| 16,384 | 1,707.49 | 220.78 |
| 32,768 | 1,366.11 | 206.37 |
| 65,536 | 976.97 | 177.97 |
| 131,072 | 637.00 | 143.90 |
| 261,312 | 376.25 | 103.12 |
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 | 1,907.55 | 290.68 | 145.34 | 679.08 |
| 3 | 8,192 | 1,751.47 | 326.24 | 108.75 | 793.94 |
| 4 | 8,192 | 1,629.37 | 341.78 | 85.45 | 858.07 |
| 5 | 8,192 | 1,567.76 | 343.97 | 68.79 | 888.08 |
| 6 | 8,192 | 1,566.95 | 356.89 | 59.48 | 938.50 |
| 7 | 8,192 | 1,535.07 | 350.47 | 50.07 | 940.53 |
| 8 | 8,192 | 1,556.47 | 364.42 | 45.55 | 989.97 |
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-3-c1
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-3-cfb69116b1
- 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 3
- 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 2,619.87 t/s (min 2,598.88, max 2,636.51, SD 15.67, n=3); TG mean 243.73 t/s (min 242.17, max 244.71, SD 1.12, n=3); All mean 600.44 t/s (min 597.97, max 601.83, SD 1.75, n=3) · KV 2,048, 1 stream: PP mean 2,134.34 t/s (min 2,092.94, max 2,192.57, SD 42.38, n=3); TG mean 238.32 t/s (min 237.19, max 239.97, SD 1.19, n=3); All mean 569.61 t/s (min 565.25, max 575.46, SD 4.30, n=3) · KV 4,096, 1 stream: PP mean 2,132.94 t/s (min 2,127.10, max 2,135.98, SD 4.13, n=3); TG mean 237.02 t/s (min 236.65, max 237.38, SD 0.30, n=3); All mean 567.07 t/s (min 565.69, max 568.12, SD 1.02, n=3) · KV 8,192, 1 stream: PP mean 2,003.78 t/s (min 1,969.27, max 2,023.45, SD 24.48, n=3); TG mean 231.90 t/s (min 230.38, max 232.83, SD 1.09, n=3); All mean 550.90 t/s (min 546.35, max 553.39, SD 3.22, n=3) · KV 16,384, 1 stream: PP mean 1,707.49 t/s (min 1,667.65, max 1,733.14, SD 28.55, n=3); TG mean 220.78 t/s (min 219.72, max 222.77, SD 1.41, n=3); All mean 512.77 t/s (min 508.39, max 518.42, SD 4.19, n=3) · KV 32,768, 1 stream: PP mean 1,366.11 t/s (min 1,317.41, max 1,396.19, SD 34.76, n=3); TG mean 206.37 t/s (min 205.83, max 207.20, SD 0.59, n=3); All mean 463.36 t/s (min 458.63, max 466.89, SD 3.48, n=3) · KV 65,536, 1 stream: PP mean 976.97 t/s (min 955.67, max 992.30, SD 15.54, n=3); TG mean 177.97 t/s (min 176.55, max 180.56, SD 1.83, n=3); All mean 380.38 t/s (min 376.88, max 385.42, SD 3.65, n=3) · KV 131,072, 1 stream: PP mean 637.00 t/s (min 630.65, max 641.37, SD 4.60, n=3); TG mean 143.90 t/s (min 143.68, max 144.23, SD 0.24, n=3); All mean 288.91 t/s (min 287.27, max 289.98, SD 1.18, n=3) · KV 261,312, 1 stream: PP mean 376.25 t/s (min 375.25, max 378.18, SD 1.36, n=3); TG mean 103.12 t/s (min 102.67, max 103.47, SD 0.33, n=3); All mean 194.08 t/s (min 193.61, max 194.35, SD 0.34, 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-3, 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-3-c2
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-3-cfb69116b1
- 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 3
- 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 1,907.55 t/s (min 1,752.84, max 1,985.86, SD 109.40, n=3); TG mean 290.68 t/s (min 278.49, max 297.89, SD 8.67, n=3); All mean 679.08 t/s (min 660.98, max 700.06, SD 16.08, n=3)
- Measurement outcomes
- 8,192 tokens · 2 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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-3-c3
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-3-cfb69116b1
- 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 3
- 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 1,751.47 t/s (min 1,695.59, max 1,784.21, SD 39.71, n=3); TG mean 326.24 t/s (min 320.30, max 331.04, SD 4.45, n=3); All mean 793.94 t/s (min 782.82, max 800.40, SD 7.90, n=3)
- Measurement outcomes
- 8,192 tokens · 3 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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-3-c4
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-3-cfb69116b1
- 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 3
- 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 1,629.37 t/s (min 1,614.26, max 1,652.34, SD 16.51, n=3); TG mean 341.78 t/s (min 339.68, max 342.89, SD 1.49, n=3); All mean 858.07 t/s (min 856.06, max 861.43, SD 2.39, n=3)
- Measurement outcomes
- 8,192 tokens · 4 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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-3-c5
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-3-cfb69116b1
- 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 3
- 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 1,567.76 t/s (min 1,507.46, max 1,614.87, SD 44.83, n=3); TG mean 343.97 t/s (min 327.44, max 354.79, SD 11.87, n=3); All mean 888.08 t/s (min 850.65, max 911.11, SD 26.70, n=3)
- Measurement outcomes
- 8,192 tokens · 5 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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-3-c6
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-3-cfb69116b1
- 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 3
- 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 1,566.95 t/s (min 1,548.42, max 1,597.84, SD 21.99, n=3); TG mean 356.89 t/s (min 356.74, max 357.10, SD 0.15, n=3); All mean 938.50 t/s (min 935.77, max 940.86, SD 2.10, n=3)
- Measurement outcomes
- 8,192 tokens · 6 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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-3-c7
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-3-cfb69116b1
- 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 3
- 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 1,535.07 t/s (min 1,506.75, max 1,557.10, SD 21.03, n=3); TG mean 350.47 t/s (min 331.23, max 361.07, SD 13.63, n=3); All mean 940.53 t/s (min 891.43, max 968.00, SD 34.80, n=3)
- Measurement outcomes
- 8,192 tokens · 7 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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-3-c8
- Page entry ID
- qwen3-6-35b-a3b-ud-q3-k-xl-single-gpu-20261016-mtp-3-cfb69116b1
- 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 3
- 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 1,556.47 t/s (min 1,511.65, max 1,579.90, SD 31.70, n=3); TG mean 364.42 t/s (min 361.61, max 365.97, SD 1.99, n=3); All mean 989.97 t/s (min 978.17, max 997.10, SD 8.40, n=3)
- Measurement outcomes
- 8,192 tokens · 8 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, 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