Gemma 4 26B-A4B
google/gemma-4-26B-A4B-it
Google · 4 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-Q4_K_XL · MTP 1
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 3,692.91 | 203.68 |
| 2,048 | 2,022.56 | 197.55 |
| 4,096 | 1,997.27 | 192.68 |
| 8,192 | 1,915.67 | 186.44 |
| 16,384 | 1,773.54 | 172.82 |
| 32,768 | 1,576.78 | 152.33 |
| 65,536 | 1,285.08 | 123.30 |
| 131,072 | 902.86 | 90.11 |
| 261,312 | 593.28 | 58.35 |
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 | 1,678.65 | 234.72 | 117.36 | 562.04 |
| 3 | 8,192 | 1,607.17 | 248.62 | 82.87 | 620.52 |
| 4 | 8,192 | 1,538.07 | 244.23 | 61.06 | 648.00 |
| 5 | 8,192 | 1,416.55 | 269.28 | 53.86 | 709.10 |
| 6 | 8,192 | 1,435.87 | 278.75 | 46.46 | 748.87 |
| 7 | 8,192 | 1,417.91 | 283.18 | 40.45 | 770.07 |
| 8 | 8,192 | 1,372.50 | 286.45 | 35.81 | 785.25 |
Technical record and provenance
Source 1 · 2026-10-14
- Source run ID
- pub-20261016-trpro-gemma26-q4-k-xl-context-2026-10-14-single-gpu-isolated-mtp-1-c1
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-1-b52ffd286b
- Entry sweep
- context
- Raw entry label
- Gemma 4 26B-A4B UD-Q4_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-14
- 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 1
- 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 3,692.91 t/s (min 3,655.26, max 3,726.29, SD 29.15, n=3); TG mean 203.68 t/s (min 203.29, max 204.33, SD 0.47, n=3); All mean 537.81 t/s (min 537.04, max 538.92, SD 0.80, n=3) · KV 2,048, 1 stream: PP mean 2,022.56 t/s (min 2,016.63, max 2,033.04, SD 7.43, n=3); TG mean 197.55 t/s (min 197.40, max 197.76, SD 0.15, n=3); All mean 485.49 t/s (min 484.95, max 486.37, SD 0.63, n=3) · KV 4,096, 1 stream: PP mean 1,997.27 t/s (min 1,974.49, max 2,011.15, SD 16.23, n=3); TG mean 192.68 t/s (min 192.04, max 193.21, SD 0.49, n=3); All mean 474.36 t/s (min 472.55, max 475.60, SD 1.31, n=3) · KV 8,192, 1 stream: PP mean 1,915.67 t/s (min 1,910.93, max 1,920.33, SD 3.84, n=3); TG mean 186.44 t/s (min 185.91, max 186.76, SD 0.38, n=3); All mean 458.17 t/s (min 457.31, max 458.72, SD 0.62, n=3) · KV 16,384, 1 stream: PP mean 1,773.54 t/s (min 1,732.91, max 1,812.75, SD 32.61, n=3); TG mean 172.82 t/s (min 171.43, max 174.04, SD 1.08, n=3); All mean 424.73 t/s (min 422.08, max 428.52, SD 2.75, n=3) · KV 32,768, 1 stream: PP mean 1,576.78 t/s (min 1,575.93, max 1,578.29, SD 1.07, n=3); TG mean 152.33 t/s (min 151.55, max 153.00, SD 0.59, n=3); All mean 375.02 t/s (min 373.45, max 376.38, SD 1.21, n=3) · KV 65,536, 1 stream: PP mean 1,285.08 t/s (min 1,280.90, max 1,290.83, SD 4.20, n=3); TG mean 123.30 t/s (min 121.70, max 124.17, SD 1.13, n=3); All mean 303.68 t/s (min 300.74, max 305.27, SD 2.08, n=3) · KV 131,072, 1 stream: PP mean 902.86 t/s (min 885.67, max 925.08, SD 16.48, n=3); TG mean 90.11 t/s (min 89.78, max 90.64, SD 0.38, n=3); All mean 220.64 t/s (min 219.39, max 222.50, SD 1.34, n=3) · KV 261,312, 1 stream: PP mean 593.28 t/s (min 588.31, max 601.76, SD 6.02, n=3); TG mean 58.35 t/s (min 57.91, max 58.73, SD 0.34, n=3); All mean 143.25 t/s (min 142.13, max 144.50, SD 0.97, 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-1, 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/gemma26-q4-k-xl-context-isolated
Source 2 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma26-q4-k-xl-concurrency-2026-10-14-single-gpu-isolated-mtp-1-c2
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-1-b52ffd286b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 26B-A4B UD-Q4_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 1
- 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 1,678.65 t/s (min 1,570.27, max 1,826.93, SD 108.51, n=3); TG mean 234.72 t/s (min 233.03, max 236.46, SD 1.40, n=3); All mean 562.04 t/s (min 554.12, max 570.12, SD 6.54, n=3)
- Measurement outcomes
- 8,192 tokens · 2 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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/2026-10-14-approved-models-single-gpu/gemma26-q4-k-xl-concurrency-isolated
Source 3 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma26-q4-k-xl-concurrency-2026-10-14-single-gpu-isolated-mtp-1-c3
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-1-b52ffd286b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 26B-A4B UD-Q4_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 1
- 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 1,607.17 t/s (min 1,562.95, max 1,669.55, SD 45.38, n=3); TG mean 248.62 t/s (min 247.53, max 249.37, SD 0.79, n=3); All mean 620.52 t/s (min 614.82, max 626.72, SD 4.87, n=3)
- Measurement outcomes
- 8,192 tokens · 3 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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/2026-10-14-approved-models-single-gpu/gemma26-q4-k-xl-concurrency-isolated
Source 4 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma26-q4-k-xl-concurrency-2026-10-14-single-gpu-isolated-mtp-1-c4
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-1-b52ffd286b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 26B-A4B UD-Q4_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 1
- 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 1,538.07 t/s (min 1,364.52, max 1,627.33, SD 122.74, n=3); TG mean 244.23 t/s (min 240.91, max 246.15, SD 2.36, n=3); All mean 648.00 t/s (min 623.84, max 660.88, SD 17.09, n=3)
- Measurement outcomes
- 8,192 tokens · 4 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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/2026-10-14-approved-models-single-gpu/gemma26-q4-k-xl-concurrency-isolated
Source 5 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma26-q4-k-xl-concurrency-2026-10-14-single-gpu-isolated-mtp-1-c5
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-1-b52ffd286b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 26B-A4B UD-Q4_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 1
- 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 1,416.55 t/s (min 1,381.13, max 1,451.88, SD 28.89, n=3); TG mean 269.28 t/s (min 267.20, max 270.51, SD 1.48, n=3); All mean 709.10 t/s (min 702.82, max 714.89, SD 4.94, n=3)
- Measurement outcomes
- 8,192 tokens · 5 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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/2026-10-14-approved-models-single-gpu/gemma26-q4-k-xl-concurrency-isolated
Source 6 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma26-q4-k-xl-concurrency-2026-10-14-single-gpu-isolated-mtp-1-c6
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-1-b52ffd286b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 26B-A4B UD-Q4_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 1
- 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 1,435.87 t/s (min 1,432.22, max 1,439.47, SD 2.96, n=3); TG mean 278.75 t/s (min 278.45, max 278.93, SD 0.22, n=3); All mean 748.87 t/s (min 748.26, max 749.33, SD 0.45, n=3)
- Measurement outcomes
- 8,192 tokens · 6 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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/2026-10-14-approved-models-single-gpu/gemma26-q4-k-xl-concurrency-isolated
Source 7 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma26-q4-k-xl-concurrency-2026-10-14-single-gpu-isolated-mtp-1-c7
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-1-b52ffd286b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 26B-A4B UD-Q4_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 1
- 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 1,417.91 t/s (min 1,410.59, max 1,422.56, SD 5.24, n=3); TG mean 283.18 t/s (min 282.16, max 283.72, SD 0.72, n=3); All mean 770.07 t/s (min 768.36, max 771.97, SD 1.48, n=3)
- Measurement outcomes
- 8,192 tokens · 7 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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/2026-10-14-approved-models-single-gpu/gemma26-q4-k-xl-concurrency-isolated
Source 8 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma26-q4-k-xl-concurrency-2026-10-14-single-gpu-isolated-mtp-1-c8
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-1-b52ffd286b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 26B-A4B UD-Q4_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 1
- 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 1,372.50 t/s (min 1,339.42, max 1,408.68, SD 28.36, n=3); TG mean 286.45 t/s (min 283.92, max 289.60, SD 2.36, n=3); All mean 785.25 t/s (min 777.05, max 796.22, SD 8.07, n=3)
- Measurement outcomes
- 8,192 tokens · 8 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-1, 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/2026-10-14-approved-models-single-gpu/gemma26-q4-k-xl-concurrency-isolated