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 3
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 3,561.72 | 235.57 |
| 2,048 | 1,925.30 | 253.77 |
| 4,096 | 1,974.55 | 248.92 |
| 8,192 | 1,874.83 | 239.79 |
| 16,384 | 1,745.69 | 222.47 |
| 32,768 | 1,548.86 | 194.76 |
| 65,536 | 1,294.24 | 160.52 |
| 131,072 | 930.48 | 116.07 |
| 261,312 | 597.91 | 73.83 |
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 | 1,763.27 | 303.43 | 151.72 | 693.49 |
| 3 | 8,192 | 1,576.49 | 346.06 | 115.35 | 813.25 |
| 4 | 8,192 | 1,527.36 | 350.82 | 87.71 | 871.56 |
| 5 | 8,192 | 1,493.03 | 359.96 | 71.99 | 914.58 |
| 6 | 8,192 | 1,464.73 | 357.08 | 59.51 | 931.53 |
| 7 | 8,192 | 1,449.44 | 357.75 | 51.11 | 950.64 |
| 8 | 8,192 | 1,425.15 | 362.42 | 45.30 | 973.58 |
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-3-c1
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-3-ae5c254273
- 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 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 3,561.72 t/s (min 3,521.93, max 3,619.40, SD 41.75, n=3); TG mean 235.57 t/s (min 235.11, max 236.18, SD 0.45, n=3); All mean 607.96 t/s (min 607.39, max 608.25, SD 0.40, n=3) · KV 2,048, 1 stream: PP mean 1,925.30 t/s (min 1,912.80, max 1,945.14, SD 14.18, n=3); TG mean 253.77 t/s (min 253.53, max 254.06, SD 0.22, n=3); All mean 587.52 t/s (min 586.57, max 589.37, SD 1.31, n=3) · KV 4,096, 1 stream: PP mean 1,974.55 t/s (min 1,926.17, max 2,006.43, SD 34.78, n=3); TG mean 248.92 t/s (min 247.07, max 250.01, SD 1.31, n=3); All mean 581.07 t/s (min 574.78, max 584.23, SD 4.45, n=3) · KV 8,192, 1 stream: PP mean 1,874.83 t/s (min 1,811.15, max 1,914.61, SD 45.49, n=3); TG mean 239.79 t/s (min 237.62, max 241.84, SD 1.73, n=3); All mean 558.08 t/s (min 549.94, max 564.44, SD 6.05, n=3) · KV 16,384, 1 stream: PP mean 1,745.69 t/s (min 1,677.23, max 1,823.39, SD 60.02, n=3); TG mean 222.47 t/s (min 219.86, max 224.13, SD 1.87, n=3); All mean 517.90 t/s (min 509.00, max 525.36, SD 6.76, n=3) · KV 32,768, 1 stream: PP mean 1,548.86 t/s (min 1,522.26, max 1,589.01, SD 28.89, n=3); TG mean 194.76 t/s (min 193.09, max 197.09, SD 1.70, n=3); All mean 454.79 t/s (min 450.91, max 461.37, SD 4.67, n=3) · KV 65,536, 1 stream: PP mean 1,294.24 t/s (min 1,281.82, max 1,303.40, SD 9.11, n=3); TG mean 160.52 t/s (min 159.94, max 161.48, SD 0.68, n=3); All mean 375.80 t/s (min 374.12, max 378.22, SD 1.76, n=3) · KV 131,072, 1 stream: PP mean 930.48 t/s (min 925.91, max 935.51, SD 3.93, n=3); TG mean 116.07 t/s (min 115.81, max 116.22, SD 0.18, n=3); All mean 271.56 t/s (min 271.16, max 271.80, SD 0.28, n=3) · KV 261,312, 1 stream: PP mean 597.91 t/s (min 590.40, max 602.07, SD 5.32, n=3); TG mean 73.83 t/s (min 72.86, max 75.13, SD 0.96, n=3); All mean 173.23 t/s (min 170.95, max 175.68, SD 1.94, 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/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-3-c2
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-3-ae5c254273
- 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 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 1,763.27 t/s (min 1,539.62, max 1,893.68, SD 158.87, n=3); TG mean 303.43 t/s (min 297.50, max 306.72, SD 4.21, n=3); All mean 693.49 t/s (min 660.75, max 711.55, SD 23.19, 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/2026-10-14-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-3-c3
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-3-ae5c254273
- 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 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 1,576.49 t/s (min 1,522.47, max 1,682.77, SD 75.15, n=3); TG mean 346.06 t/s (min 338.58, max 352.16, SD 5.63, n=3); All mean 813.25 t/s (min 799.98, max 833.89, SD 14.79, 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/2026-10-14-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-3-c4
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-3-ae5c254273
- 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 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 1,527.36 t/s (min 1,451.94, max 1,621.40, SD 70.42, n=3); TG mean 350.82 t/s (min 348.57, max 352.95, SD 1.79, n=3); All mean 871.56 t/s (min 850.90, max 900.04, SD 20.81, 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/2026-10-14-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-3-c5
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-3-ae5c254273
- 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 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 1,493.03 t/s (min 1,486.95, max 1,496.74, SD 4.33, n=3); TG mean 359.96 t/s (min 358.17, max 360.88, SD 1.27, n=3); All mean 914.58 t/s (min 910.53, max 916.65, SD 2.87, 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/2026-10-14-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-3-c6
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-3-ae5c254273
- 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 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 1,464.73 t/s (min 1,453.70, max 1,471.23, SD 7.84, n=3); TG mean 357.08 t/s (min 356.34, max 357.92, SD 0.65, n=3); All mean 931.53 t/s (min 929.17, max 933.93, SD 1.94, 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/2026-10-14-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-3-c7
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-3-ae5c254273
- 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 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 1,449.44 t/s (min 1,446.51, max 1,453.89, SD 3.20, n=3); TG mean 357.75 t/s (min 356.28, max 359.36, SD 1.26, n=3); All mean 950.64 t/s (min 947.54, max 953.73, SD 2.53, 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/2026-10-14-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-3-c8
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-mtp-3-ae5c254273
- 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 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 1,425.15 t/s (min 1,416.08, max 1,431.12, SD 6.52, n=3); TG mean 362.42 t/s (min 361.12, max 363.35, SD 0.95, n=3); All mean 973.58 t/s (min 971.28, max 976.99, SD 2.46, 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/2026-10-14-approved-models-single-gpu/gemma26-q4-k-xl-concurrency-isolated