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-IQ4_XS · MTP 2
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 2,597.06 | 220.53 |
| 2,048 | 2,210.48 | 219.65 |
| 4,096 | 2,101.66 | 215.07 |
| 8,192 | 2,035.81 | 209.80 |
| 16,384 | 1,671.37 | 199.40 |
| 32,768 | 1,362.33 | 182.42 |
| 65,536 | 980.20 | 158.81 |
| 131,072 | 612.28 | 123.94 |
| 261,312 | 370.85 | 86.66 |
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 | 1,816.74 | 269.51 | 134.76 | 634.46 |
| 3 | 8,192 | 1,683.43 | 278.65 | 92.88 | 687.69 |
| 4 | 8,192 | 1,508.10 | 290.77 | 72.69 | 745.82 |
| 5 | 8,192 | 1,516.68 | 307.37 | 61.47 | 805.82 |
| 6 | 8,192 | 1,471.30 | 310.51 | 51.75 | 828.65 |
| 7 | 8,192 | 1,455.73 | 318.90 | 45.56 | 862.11 |
| 8 | 8,192 | 1,461.34 | 325.09 | 40.64 | 890.10 |
Technical record and provenance
Source 1 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-iq4-xs-context-2026-10-14-single-gpu-isolated-mtp-2-c1
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-2-9cad5907fd
- Entry sweep
- context
- Raw entry label
- Qwen3.6 35B-A3B UD-IQ4_XS — 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 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 2,597.06 t/s (min 2,556.38, max 2,617.81, SD 28.77, n=3); TG mean 220.53 t/s (min 219.94, max 221.15, SD 0.50, n=3); All mean 552.29 t/s (min 551.01, max 554.59, SD 1.64, n=3) · KV 2,048, 1 stream: PP mean 2,210.48 t/s (min 2,206.23, max 2,213.55, SD 3.10, n=3); TG mean 219.65 t/s (min 219.63, max 219.67, SD 0.02, n=3); All mean 536.92 t/s (min 536.62, max 537.10, SD 0.21, n=3) · KV 4,096, 1 stream: PP mean 2,101.66 t/s (min 2,079.12, max 2,146.36, SD 31.61, n=3); TG mean 215.07 t/s (min 213.11, max 216.27, SD 1.40, n=3); All mean 523.28 t/s (min 518.32, max 527.37, SD 3.74, n=3) · KV 8,192, 1 stream: PP mean 2,035.81 t/s (min 2,022.04, max 2,051.08, SD 11.91, n=3); TG mean 209.80 t/s (min 208.50, max 210.59, SD 0.93, n=3); All mean 509.53 t/s (min 507.80, max 510.96, SD 1.31, n=3) · KV 16,384, 1 stream: PP mean 1,671.37 t/s (min 1,639.28, max 1,724.09, SD 37.57, n=3); TG mean 199.40 t/s (min 197.79, max 201.11, SD 1.35, n=3); All mean 471.83 t/s (min 467.10, max 477.90, SD 4.51, n=3) · KV 32,768, 1 stream: PP mean 1,362.33 t/s (min 1,303.28, max 1,395.31, SD 41.85, n=3); TG mean 182.42 t/s (min 180.56, max 183.59, SD 1.33, n=3); All mean 421.46 t/s (min 419.04, max 425.41, SD 2.82, n=3) · KV 65,536, 1 stream: PP mean 980.20 t/s (min 943.42, max 999.54, SD 26.01, n=3); TG mean 158.81 t/s (min 158.23, max 159.29, SD 0.44, n=3); All mean 351.24 t/s (min 347.09, max 353.59, SD 2.94, n=3) · KV 131,072, 1 stream: PP mean 612.28 t/s (min 598.27, max 636.28, SD 17.05, n=3); TG mean 123.94 t/s (min 123.58, max 124.41, SD 0.35, n=3); All mean 258.20 t/s (min 256.10, max 260.91, SD 2.01, n=3) · KV 261,312, 1 stream: PP mean 370.85 t/s (min 364.29, max 374.78, SD 4.67, n=3); TG mean 86.66 t/s (min 86.09, max 86.99, SD 0.40, n=3); All mean 172.83 t/s (min 172.22, max 173.98, SD 0.81, 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/qwen36-iq4-xs-context-isolated
Source 2 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-iq4-xs-concurrency-2026-10-14-single-gpu-isolated-mtp-2-c2
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-2-9cad5907fd
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-IQ4_XS — 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 1,816.74 t/s (min 1,745.48, max 1,916.29, SD 72.55, n=3); TG mean 269.51 t/s (min 267.00, max 271.08, SD 1.79, n=3); All mean 634.46 t/s (min 625.13, max 643.20, SD 7.39, 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/qwen36-iq4-xs-concurrency-isolated
Source 3 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-iq4-xs-concurrency-2026-10-14-single-gpu-isolated-mtp-2-c3
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-2-9cad5907fd
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-IQ4_XS — 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 1,683.43 t/s (min 1,633.99, max 1,741.75, SD 44.44, n=3); TG mean 278.65 t/s (min 278.02, max 279.39, SD 0.56, n=3); All mean 687.69 t/s (min 682.49, max 693.01, SD 4.30, 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/qwen36-iq4-xs-concurrency-isolated
Source 4 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-iq4-xs-concurrency-2026-10-14-single-gpu-isolated-mtp-2-c4
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-2-9cad5907fd
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-IQ4_XS — 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 1,508.10 t/s (min 1,450.63, max 1,557.91, SD 44.13, n=3); TG mean 290.77 t/s (min 288.24, max 292.07, SD 1.79, n=3); All mean 745.82 t/s (min 734.95, max 755.11, SD 8.31, 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/qwen36-iq4-xs-concurrency-isolated
Source 5 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-iq4-xs-concurrency-2026-10-14-single-gpu-isolated-mtp-2-c5
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-2-9cad5907fd
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-IQ4_XS — 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 1,516.68 t/s (min 1,496.99, max 1,545.83, SD 21.03, n=3); TG mean 307.37 t/s (min 305.76, max 309.34, SD 1.48, n=3); All mean 805.82 t/s (min 801.33, max 811.87, SD 4.45, 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/qwen36-iq4-xs-concurrency-isolated
Source 6 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-iq4-xs-concurrency-2026-10-14-single-gpu-isolated-mtp-2-c6
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-2-9cad5907fd
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-IQ4_XS — 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 1,471.30 t/s (min 1,435.16, max 1,489.64, SD 25.56, n=3); TG mean 310.51 t/s (min 308.66, max 311.54, SD 1.31, n=3); All mean 828.65 t/s (min 820.85, max 833.27, SD 5.54, 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/qwen36-iq4-xs-concurrency-isolated
Source 7 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-iq4-xs-concurrency-2026-10-14-single-gpu-isolated-mtp-2-c7
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-2-9cad5907fd
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-IQ4_XS — 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 1,455.73 t/s (min 1,450.16, max 1,465.12, SD 6.68, n=3); TG mean 318.90 t/s (min 318.40, max 319.16, SD 0.35, n=3); All mean 862.11 t/s (min 860.39, max 864.68, SD 1.85, 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/qwen36-iq4-xs-concurrency-isolated
Source 8 · 2026-10-15
- Source run ID
- pub-20261016-trpro-qwen36-iq4-xs-concurrency-2026-10-14-single-gpu-isolated-mtp-2-c8
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-2-9cad5907fd
- Entry sweep
- concurrency
- Raw entry label
- Qwen3.6 35B-A3B UD-IQ4_XS — 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 1,461.34 t/s (min 1,421.29, max 1,502.84, SD 33.31, n=3); TG mean 325.09 t/s (min 322.79, max 328.77, SD 2.63, n=3); All mean 890.10 t/s (min 880.40, max 902.11, SD 9.01, 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/qwen36-iq4-xs-concurrency-isolated