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 1
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 2,631.07 | 186.23 |
| 2,048 | 2,188.06 | 182.08 |
| 4,096 | 2,133.54 | 179.11 |
| 8,192 | 2,033.38 | 173.91 |
| 16,384 | 1,695.48 | 164.75 |
| 32,768 | 1,390.95 | 149.39 |
| 65,536 | 984.21 | 125.48 |
| 131,072 | 620.51 | 95.91 |
| 261,312 | 369.35 | 65.50 |
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 | 1,836.59 | 228.98 | 114.49 | 557.75 |
| 3 | 8,192 | 1,682.64 | 241.25 | 80.42 | 609.88 |
| 4 | 8,192 | 1,538.36 | 239.40 | 59.85 | 630.58 |
| 5 | 8,192 | 1,522.89 | 250.68 | 50.14 | 672.12 |
| 6 | 8,192 | 1,528.02 | 263.05 | 43.84 | 715.21 |
| 7 | 8,192 | 1,510.92 | 271.39 | 38.77 | 745.05 |
| 8 | 8,192 | 1,530.68 | 276.04 | 34.51 | 768.09 |
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-1-c1
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-1-5e8d0e448d
- 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 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 2,631.07 t/s (min 2,606.70, max 2,647.77, SD 17.62, n=3); TG mean 186.23 t/s (min 186.10, max 186.33, SD 0.10, n=3); All mean 479.51 t/s (min 479.23, max 479.99, SD 0.34, n=3) · KV 2,048, 1 stream: PP mean 2,188.06 t/s (min 2,126.25, max 2,227.27, SD 44.23, n=3); TG mean 182.08 t/s (min 181.89, max 182.29, SD 0.16, n=3); All mean 459.04 t/s (min 457.13, max 460.31, SD 1.38, n=3) · KV 4,096, 1 stream: PP mean 2,133.54 t/s (min 2,124.01, max 2,143.87, SD 8.13, n=3); TG mean 179.11 t/s (min 178.78, max 179.77, SD 0.47, n=3); All mean 451.10 t/s (min 450.15, max 452.71, SD 1.15, n=3) · KV 8,192, 1 stream: PP mean 2,033.38 t/s (min 2,026.21, max 2,038.39, SD 5.20, n=3); TG mean 173.91 t/s (min 173.02, max 174.74, SD 0.70, n=3); All mean 436.52 t/s (min 434.30, max 438.18, SD 1.63, n=3) · KV 16,384, 1 stream: PP mean 1,695.48 t/s (min 1,636.85, max 1,731.06, SD 41.77, n=3); TG mean 164.75 t/s (min 164.40, max 165.06, SD 0.27, n=3); All mean 405.70 t/s (min 402.48, max 407.80, SD 2.31, n=3) · KV 32,768, 1 stream: PP mean 1,390.95 t/s (min 1,382.46, max 1,396.09, SD 6.05, n=3); TG mean 149.39 t/s (min 149.10, max 149.66, SD 0.23, n=3); All mean 361.47 t/s (min 360.93, max 362.10, SD 0.48, n=3) · KV 65,536, 1 stream: PP mean 984.21 t/s (min 960.07, max 997.22, SD 17.09, n=3); TG mean 125.48 t/s (min 125.07, max 125.75, SD 0.30, n=3); All mean 294.00 t/s (min 291.81, max 295.35, SD 1.56, n=3) · KV 131,072, 1 stream: PP mean 620.51 t/s (min 591.76, max 637.48, SD 20.44, n=3); TG mean 95.91 t/s (min 95.75, max 96.19, SD 0.20, n=3); All mean 215.26 t/s (min 212.87, max 216.50, SD 1.69, n=3) · KV 261,312, 1 stream: PP mean 369.35 t/s (min 361.46, max 375.33, SD 5.82, n=3); TG mean 65.50 t/s (min 65.15, max 65.70, SD 0.25, n=3); All mean 142.14 t/s (min 141.56, max 143.01, SD 0.63, 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/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-1-c2
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-1-5e8d0e448d
- 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 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 1,836.59 t/s (min 1,724.45, max 1,983.72, SD 108.70, n=3); TG mean 228.98 t/s (min 227.94, max 229.69, SD 0.75, n=3); All mean 557.75 t/s (min 549.73, max 567.03, SD 7.12, 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/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-1-c3
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-1-5e8d0e448d
- 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 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 1,682.64 t/s (min 1,627.75, max 1,747.16, SD 49.22, n=3); TG mean 241.25 t/s (min 240.68, max 242.29, SD 0.73, n=3); All mean 609.88 t/s (min 605.52, max 616.89, SD 5.01, 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/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-1-c4
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-1-5e8d0e448d
- 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 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 1,538.36 t/s (min 1,525.44, max 1,548.95, SD 9.74, n=3); TG mean 239.40 t/s (min 237.48, max 240.48, SD 1.36, n=3); All mean 630.58 t/s (min 628.26, max 631.91, SD 1.65, 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/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-1-c5
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-1-5e8d0e448d
- 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 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 1,522.89 t/s (min 1,473.05, max 1,549.66, SD 35.28, n=3); TG mean 250.68 t/s (min 249.67, max 251.60, SD 0.79, n=3); All mean 672.12 t/s (min 667.23, max 674.78, SD 3.46, 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/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-1-c6
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-1-5e8d0e448d
- 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 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 1,528.02 t/s (min 1,520.35, max 1,532.53, SD 5.45, n=3); TG mean 263.05 t/s (min 261.52, max 264.04, SD 1.10, n=3); All mean 715.21 t/s (min 711.34, max 718.03, SD 2.83, 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/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-1-c7
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-1-5e8d0e448d
- 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 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 1,510.92 t/s (min 1,490.37, max 1,531.78, SD 16.91, n=3); TG mean 271.39 t/s (min 268.43, max 272.89, SD 2.09, n=3); All mean 745.05 t/s (min 737.38, max 748.96, SD 5.42, 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/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-1-c8
- Page entry ID
- qwen3-6-35b-a3b-ud-iq4-xs-single-gpu-20261016-mtp-1-5e8d0e448d
- 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 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 1,530.68 t/s (min 1,465.10, max 1,567.99, SD 46.51, n=3); TG mean 276.04 t/s (min 274.51, max 278.12, SD 1.52, n=3); All mean 768.09 t/s (min 759.64, max 774.95, SD 6.35, 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/qwen36-iq4-xs-concurrency-isolated