Report · resources

KTS-1 standard

Qwen3.6-35B-A3B · NVFP4 variant · NVIDIA NGC vLLM 0.29

Measured results / KTS-1 standard / Run 4bf8d4f0df2d4dce828cbfd6

Measured 2026-10-18 20:59 UTC · Record date 2026-10-18

Measured.

KTS-1 adaptive capacity

Measured · 5 people × 40K new (distinct inputs)

Measured on one configuration (variant 415588da), 3 repeats. Per-person values for each repeat are in this run’s OLIR JSON.

Repeats · 3
Recorded repeats for this run
RepeatEach person · medianMachine totalFirst word · medianSlowest first wordOverlapping decodesPower · medianEnergyTokens per kWhService barRun
1
50.24 tok/s
206.70 tok/s
41.44 s
41.45 s
5
56.2 W
0.00079 kWh
587,465
met
2
49.71 tok/s
140.17 tok/s
41.21 s
41.21 s
5
58.8 W
0.00082 kWh
580,857
met
3
49.45 tok/s
146.79 tok/s
40.70 s
40.70 s
5
60.5 W
0.00084 kWh
545,022
met
Method revision
KTS-1/adaptive-capacity-v2
Protocol SHA-256
25fe3b0656961091ac1dfcc41df0b8cce189498814bf6c134c310f95ac15d1b0
Implementation SHA-256
2de3c503cd76f5fc9ce6d5148760c219ddb52ee5cfbf5d728c3697037536d047
Cell outcome SHA-256
7e9705d9e6e9c55c1b3f235a3a2b30ae6b07a05884e0a62e314ec4c8b8d27551
Attempts · 1
Attempts recorded for this cell
RoleClassificationCapacity boundaryConfigurationAttempt SHA-256Run
primary
measured
none
415588da
c09dec415600b431ce80e61c902526a5f9e5b4ec745064331dee24ca1697d51e

Runs on this setup

Earlier runs keep their own engine versions and outcomes. Workload shapes stay separate.

History and measured sweep
History of this setup, newest first
DateCell or blockEach person · medianResultRun
2026-10-19
5 people · 250K · already read
35.94 tok/s
Measured
2026-10-19
2 people · 250K · already read
49.98 tok/s
Measured
2026-10-19
1 people · 250K · already read
61.49 tok/s
Measured
2026-10-19
5 people · 250K · new
18.82 tok/s
Measured
2026-10-19
2 people · 250K · new
30.96 tok/s
Measured
2026-10-19
1 people · 250K · new
61.96 tok/s
Measured
2026-10-18
5 people · 128K · new
25.02 tok/s
Measured
2026-10-18
2 people · 128K · new
45.22 tok/s
Measured
2026-10-18
1 people · 128K · new
73.46 tok/s
Measured
2026-10-18
8 people · 64K · new
6.88 tok/s
Measured
2026-10-18
5 people · 64K · new
44.58 tok/s
Measured
2026-10-18
2 people · 64K · new
59.94 tok/s
Measured
2026-10-18
1 people · 64K · new
88.61 tok/s
Measured
2026-10-18
8 people · 40K · new
7.62 tok/s
Measured
2026-10-18
5 people · 40K · new
49.71 tok/s
Measured
2026-10-18
2 people · 40K · new
73.60 tok/s
Measured
2026-10-18
1 people · 40K · new
93.71 tok/s
Measured
2026-10-18
20 people · 4K · new
16.74 tok/s
Measured
2026-10-18
16 people · 4K · new
23.50 tok/s
Measured
2026-10-18
8 people · 4K · new
43.17 tok/s
Measured
2026-10-18
5 people · 4K · new
65.78 tok/s
Measured
2026-10-18
2 people · 4K · new
86.44 tok/s
Measured
2026-10-18
1 people · 4K · new
104.04 tok/s
Measured

Answer checks

No answer-quality suite was graded for this run.

Method

KTS-1 adaptive-capacity-v2, owner amendment 2026-10-14. Ascending concurrency per context; one different same-model configuration confirmation of observed capacity. Limits have no numeric measurements.

Limits

Labelled lab evidence for the recorded setup and standard. Completion is coverage, not customer or general answer-quality qualification.

What ran

Exact identity

Original publisher model
Qwen3.6-35B-A3B
Publisher revision
not recorded
Variant · format · precision
NVFP4 variant · safetensors · nvfp4
Weights SHA-256
07141c2db92e47bc08777132cd1a0323faf300eab3a7d7c111bc2bf075fda050
6dea9c759a0f941cf9e1cc1501216b0e107966a66c90425050e212efbd053f02
9758875fc55e49561165f4a44342b654c123c3e25c6811b34abb83553fb1a164
Engine · version · feature
NVIDIA NGC vLLM · 0.29.0+5013de39.dev · plain
Hardware
NVIDIA GB10, 128 GB unified memory
Harness / workload
not recorded
Workload suite · version
not recorded
Harness SHA-256
not recorded
Fixture SHA-256
not recorded

Workload shape

adaptive variant sha256
415588daa437380e860555f3e55a80e6b6c9c4c8dce429d6c940f2fb436ef261
cache state
cold
People in this run
5
Configured window (tokens)
262144
Configured window (tokens)
40000
scenario
distinct

Full reproduction pack available to customers and evaluators on request

All recorded numbers

Metric definitions, scope and statistics
All aggregate numeric result fields in this run
MetricValueScopeStatisticSample countBasis / source fieldRun
median decode tps
49.71 tok/s
per person
median
not recorded
standard_result.median_decode_tps
n
3.00 count
machine total
single
not recorded
standard_result.n

Data and citation

Skilak Keep evidence, run 4bf8d4f0df2d4dce828cbfd6, measured 2026-10-18T20:59:08.748957+00:00.
https://skilak.ai/keep/resources/evidence/r/4bf8d4f0df2d4dce828cbfd6

Ready to scope it?

Start with one workload.

Tell us who uses it, what it can read, and what must stay inside. We will map the build.