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 2
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 3,664.14 | 232.25 |
| 2,048 | 1,991.65 | 229.92 |
| 4,096 | 1,975.95 | 225.42 |
| 8,192 | 1,889.43 | 218.56 |
| 16,384 | 1,811.16 | 203.73 |
| 32,768 | 1,572.62 | 179.82 |
| 65,536 | 1,279.96 | 145.41 |
| 131,072 | 910.89 | 105.43 |
| 261,312 | 595.64 | 67.62 |
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 | 1,670.55 | 280.31 | 140.16 | 644.84 |
| 3 | 8,192 | 1,602.88 | 317.67 | 105.89 | 734.78 |
| 4 | 8,192 | 1,459.18 | 307.97 | 76.99 | 783.94 |
| 5 | 8,192 | 1,392.80 | 321.03 | 64.21 | 826.90 |
| 6 | 8,192 | 1,385.45 | 320.98 | 53.50 | 848.68 |
| 7 | 8,192 | 1,383.70 | 328.23 | 46.89 | 880.31 |
| 8 | 8,192 | 1,362.25 | 329.48 | 41.19 | 893.53 |
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-2-c1
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-2-52609cfd3b
- 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 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 3,664.14 t/s (min 3,609.19, max 3,716.88, SD 43.99, n=3); TG mean 232.25 t/s (min 231.80, max 232.89, SD 0.47, n=3); All mean 602.85 t/s (min 601.81, max 603.90, SD 0.85, n=3) · KV 2,048, 1 stream: PP mean 1,991.65 t/s (min 1,927.60, max 2,024.92, SD 45.30, n=3); TG mean 229.92 t/s (min 229.63, max 230.35, SD 0.31, n=3); All mean 547.20 t/s (min 543.43, max 549.51, SD 2.69, n=3) · KV 4,096, 1 stream: PP mean 1,975.95 t/s (min 1,934.86, max 2,007.76, SD 30.48, n=3); TG mean 225.42 t/s (min 223.05, max 227.32, SD 1.78, n=3); All mean 537.41 t/s (min 534.47, max 541.45, SD 2.96, n=3) · KV 8,192, 1 stream: PP mean 1,889.43 t/s (min 1,815.91, max 1,928.50, SD 52.02, n=3); TG mean 218.56 t/s (min 217.18, max 219.68, SD 1.04, n=3); All mean 519.24 t/s (min 513.03, max 523.03, SD 4.42, n=3) · KV 16,384, 1 stream: PP mean 1,811.16 t/s (min 1,808.30, max 1,814.76, SD 2.69, n=3); TG mean 203.73 t/s (min 203.53, max 203.89, SD 0.15, n=3); All mean 486.95 t/s (min 486.40, max 487.68, SD 0.54, n=3) · KV 32,768, 1 stream: PP mean 1,572.62 t/s (min 1,563.24, max 1,585.59, SD 9.47, n=3); TG mean 179.82 t/s (min 179.14, max 180.27, SD 0.49, n=3); All mean 428.49 t/s (min 426.82, max 429.42, SD 1.18, n=3) · KV 65,536, 1 stream: PP mean 1,279.96 t/s (min 1,263.07, max 1,289.70, SD 11.99, n=3); TG mean 145.41 t/s (min 144.99, max 145.66, SD 0.30, n=3); All mean 346.85 t/s (min 345.31, max 348.00, SD 1.14, n=3) · KV 131,072, 1 stream: PP mean 910.89 t/s (min 888.55, max 922.74, SD 15.81, n=3); TG mean 105.43 t/s (min 105.11, max 105.64, SD 0.23, n=3); All mean 250.54 t/s (min 249.10, max 251.45, SD 1.03, n=3) · KV 261,312, 1 stream: PP mean 595.64 t/s (min 589.48, max 598.94, SD 4.36, n=3); TG mean 67.62 t/s (min 67.23, max 68.17, SD 0.40, n=3); All mean 161.37 t/s (min 160.75, max 162.61, SD 0.87, 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/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-2-c2
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-2-52609cfd3b
- 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 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 1,670.55 t/s (min 1,554.70, max 1,803.74, SD 102.40, n=3); TG mean 280.31 t/s (min 279.17, max 282.11, SD 1.28, n=3); All mean 644.84 t/s (min 631.59, max 657.30, SD 10.51, 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/2026-10-14-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-2-c3
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-2-52609cfd3b
- 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 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 1,602.88 t/s (min 1,486.51, max 1,732.27, SD 100.75, n=3); TG mean 317.67 t/s (min 300.44, max 344.33, SD 19.12, n=3); All mean 734.78 t/s (min 712.22, max 755.25, SD 17.63, 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/2026-10-14-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-2-c4
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-2-52609cfd3b
- 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 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 1,459.18 t/s (min 1,358.09, max 1,597.96, SD 101.49, n=3); TG mean 307.97 t/s (min 304.45, max 310.57, SD 2.58, n=3); All mean 783.94 t/s (min 763.03, max 811.65, SD 20.42, 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/2026-10-14-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-2-c5
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-2-52609cfd3b
- 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 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 1,392.80 t/s (min 1,356.68, max 1,413.05, SD 25.60, n=3); TG mean 321.03 t/s (min 318.87, max 322.43, SD 1.55, n=3); All mean 826.90 t/s (min 817.72, max 833.15, SD 6.63, 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/2026-10-14-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-2-c6
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-2-52609cfd3b
- 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 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 1,385.45 t/s (min 1,372.12, max 1,396.63, SD 10.12, n=3); TG mean 320.98 t/s (min 319.45, max 322.90, SD 1.43, n=3); All mean 848.68 t/s (min 844.48, max 853.66, SD 3.79, 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/2026-10-14-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-2-c7
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-2-52609cfd3b
- 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 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 1,383.70 t/s (min 1,378.32, max 1,388.92, SD 4.33, n=3); TG mean 328.23 t/s (min 327.19, max 329.35, SD 0.89, n=3); All mean 880.31 t/s (min 877.37, max 883.49, SD 2.50, 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/2026-10-14-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-2-c8
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-2-52609cfd3b
- 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 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 1,362.25 t/s (min 1,335.96, max 1,376.48, SD 18.62, n=3); TG mean 329.48 t/s (min 326.88, max 332.31, SD 2.23, n=3); All mean 893.53 t/s (min 884.77, max 901.83, SD 6.97, 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/2026-10-14-approved-models-single-gpu/gemma26-q4-k-xl-concurrency-isolated