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 off
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 4,565.84 | 145.11 |
| 2,048 | 2,483.23 | 137.54 |
| 4,096 | 2,495.54 | 135.51 |
| 8,192 | 2,450.46 | 130.65 |
| 16,384 | 2,245.61 | 122.06 |
| 32,768 | 1,897.74 | 107.91 |
| 65,536 | 1,509.15 | 87.46 |
| 131,072 | 1,051.49 | 63.61 |
| 261,312 | 667.15 | 41.41 |
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 | 1,926.93 | 203.95 | 101.98 | 513.27 |
| 3 | 8,192 | 1,744.81 | 223.25 | 74.42 | 579.72 |
| 4 | 8,192 | 1,604.30 | 227.45 | 56.86 | 603.06 |
| 5 | 8,192 | 1,533.33 | 231.92 | 46.38 | 631.28 |
| 6 | 8,192 | 1,453.63 | 228.31 | 38.05 | 628.28 |
| 7 | 8,192 | 1,481.50 | 232.55 | 33.22 | 647.98 |
| 8 | 8,192 | 1,458.03 | 236.59 | 29.57 | 662.19 |
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-off-c1
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-off-086d336649
- 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
- None (standard decoding)
- 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 4,565.84 t/s (min 4,478.66, max 4,619.44, SD 62.19, n=3); TG mean 145.11 t/s (min 144.54, max 145.56, SD 0.43, n=3); All mean 402.21 t/s (min 400.80, max 403.01, SD 1.00, n=3) · KV 2,048, 1 stream: PP mean 2,483.23 t/s (min 2,447.31, max 2,549.20, SD 46.71, n=3); TG mean 137.54 t/s (min 137.16, max 137.99, SD 0.34, n=3); All mean 365.49 t/s (min 364.81, max 366.16, SD 0.55, n=3) · KV 4,096, 1 stream: PP mean 2,495.54 t/s (min 2,404.68, max 2,572.81, SD 69.31, n=3); TG mean 135.51 t/s (min 135.07, max 136.04, SD 0.40, n=3); All mean 360.92 t/s (min 358.62, max 363.22, SD 1.88, n=3) · KV 8,192, 1 stream: PP mean 2,450.46 t/s (min 2,439.93, max 2,460.58, SD 8.44, n=3); TG mean 130.65 t/s (min 130.08, max 131.02, SD 0.41, n=3); All mean 348.43 t/s (min 346.95, max 349.41, SD 1.07, n=3) · KV 16,384, 1 stream: PP mean 2,245.61 t/s (min 2,206.09, max 2,271.07, SD 28.33, n=3); TG mean 122.06 t/s (min 121.86, max 122.43, SD 0.26, n=3); All mean 325.03 t/s (min 324.06, max 325.92, SD 0.76, n=3) · KV 32,768, 1 stream: PP mean 1,897.74 t/s (min 1,862.75, max 1,941.38, SD 32.68, n=3); TG mean 107.91 t/s (min 107.84, max 108.03, SD 0.08, n=3); All mean 286.07 t/s (min 285.79, max 286.55, SD 0.34, n=3) · KV 65,536, 1 stream: PP mean 1,509.15 t/s (min 1,503.11, max 1,519.68, SD 7.47, n=3); TG mean 87.46 t/s (min 87.39, max 87.53, SD 0.06, n=3); All mean 231.39 t/s (min 231.00, max 231.82, SD 0.34, n=3) · KV 131,072, 1 stream: PP mean 1,051.49 t/s (min 1,032.38, max 1,061.21, SD 13.51, n=3); TG mean 63.61 t/s (min 63.49, max 63.69, SD 0.09, n=3); All mean 167.41 t/s (min 167.16, max 167.76, SD 0.25, n=3) · KV 261,312, 1 stream: PP mean 667.15 t/s (min 664.34, max 668.69, SD 1.99, n=3); TG mean 41.41 t/s (min 41.36, max 41.44, SD 0.03, n=3); All mean 108.80 t/s (min 108.68, max 108.86, SD 0.09, 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 (off, 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-off-c2
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-off-086d336649
- 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
- None (standard decoding)
- 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 1,926.93 t/s (min 1,772.28, max 2,162.17, SD 169.07, n=3); TG mean 203.95 t/s (min 203.15, max 205.54, SD 1.12, n=3); All mean 513.27 t/s (min 505.41, max 521.69, SD 6.66, n=3)
- Measurement outcomes
- 8,192 tokens · 2 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, 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-off-c3
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-off-086d336649
- 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
- None (standard decoding)
- 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 1,744.81 t/s (min 1,698.64, max 1,828.49, SD 59.28, n=3); TG mean 223.25 t/s (min 218.86, max 225.78, SD 3.12, n=3); All mean 579.72 t/s (min 576.78, max 584.76, SD 3.58, n=3)
- Measurement outcomes
- 8,192 tokens · 3 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, 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-off-c4
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-off-086d336649
- 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
- None (standard decoding)
- 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 1,604.30 t/s (min 1,563.86, max 1,627.73, SD 28.72, n=3); TG mean 227.45 t/s (min 227.02, max 228.23, SD 0.55, n=3); All mean 603.06 t/s (min 600.14, max 605.75, SD 2.30, n=3)
- Measurement outcomes
- 8,192 tokens · 4 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, 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-off-c5
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-off-086d336649
- 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
- None (standard decoding)
- 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 1,533.33 t/s (min 1,497.56, max 1,587.76, SD 39.12, n=3); TG mean 231.92 t/s (min 231.50, max 232.26, SD 0.32, n=3); All mean 631.28 t/s (min 627.23, max 637.44, SD 4.43, n=3)
- Measurement outcomes
- 8,192 tokens · 5 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, 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-off-c6
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-off-086d336649
- 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
- None (standard decoding)
- 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 1,453.63 t/s (min 1,422.47, max 1,505.19, SD 36.72, n=3); TG mean 228.31 t/s (min 227.12, max 230.37, SD 1.46, n=3); All mean 628.28 t/s (min 624.51, max 634.94, SD 4.72, n=3)
- Measurement outcomes
- 8,192 tokens · 6 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, 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-off-c7
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-off-086d336649
- 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
- None (standard decoding)
- 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 1,481.50 t/s (min 1,448.30, max 1,498.29, SD 23.48, n=3); TG mean 232.55 t/s (min 231.37, max 233.70, SD 0.95, n=3); All mean 647.98 t/s (min 645.03, max 651.15, SD 2.50, n=3)
- Measurement outcomes
- 8,192 tokens · 7 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, 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-off-c8
- Page entry ID
- gemma-4-26b-a4b-ud-q4-k-xl-single-gpu-20261016-off-086d336649
- 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
- None (standard decoding)
- 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 1,458.03 t/s (min 1,432.39, max 1,484.87, SD 21.44, n=3); TG mean 236.59 t/s (min 235.91, max 237.77, SD 0.84, n=3); All mean 662.19 t/s (min 658.55, max 667.25, SD 3.69, n=3)
- Measurement outcomes
- 8,192 tokens · 8 streams: measured
- Notes
- Automated llama-server KV-depth sweep (off, 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