KTS-1 standard

KTS-1 adaptive capacity

Measured · 64 people × 4K new (distinct inputs)

Measured on one configuration (variant ce7a1141), 3 repeats. Per-person values for each repeat are in this run’s OLIR JSON.

Repeats · 3

Recorded repeats for this run
Repeat	Each person · median	Machine total	First word · median	Slowest first word	Overlapping decodes	Power · median	Energy	Tokens per kWh	Service bar	Run
1	4.20 tok/s	141.82 tok/s	23.30 s	40.67 s	64	12.9 W	0.00079 kWh	7,338,518	missed	Run 4301ebe8 →
2	4.30 tok/s	145.74 tok/s	22.48 s	38.99 s	64	13.2 W	0.00081 kWh	7,144,937	missed	Run 4301ebe8 →
3	4.50 tok/s	145.71 tok/s	22.50 s	39.00 s	64	13.4 W	0.00082 kWh	7,051,112	missed	Run 4301ebe8 →

Method revision: KTS-1/adaptive-capacity-v2
Protocol SHA-256: 25fe3b0656961091ac1dfcc41df0b8cce189498814bf6c134c310f95ac15d1b0
Implementation SHA-256: 2de3c503cd76f5fc9ce6d5148760c219ddb52ee5cfbf5d728c3697037536d047
Cell outcome SHA-256: d4436f6ffd490fefea0921fa550fae9bfff3daeccd3618541a6051ec97bb940a

Attempts · 2

Attempts recorded for this cell
Role	Classification	Capacity boundary	Configuration	Attempt SHA-256	Run
primary	capacity	serialization boundary	415588da	84ac08b120b0dadcf1b45cfcca0cfd8abab1c71ac1a416dea0a94a31b4c46fb4	Run 4301ebe8 →
confirmation	measured	none	ce7a1141	e16cc71812487be8206974999142c12e6c19bdfdb89b28aac1681027dece45d4	Run 4301ebe8 →

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
Date	Cell or block	Each person · median	Result	Run
2026-10-18	64 people · 4K · new	4.30 tok/s	Measured	Run 4301ebe8 →
2026-10-18	32 people · 4K · new	9.77 tok/s	Measured	Run 93f3eb0f →

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: ce7a11417103cc42fa9c8145854c46461d126b984c768acf8e243600a467a158
cache state: cold
People in this run: 64
Configured window (tokens): 262144
Configured window (tokens): 4000
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
Metric	Value	Scope	Statistic	Sample count	Basis / source field	Run
median decode tps	4.30 tok/s	per person	median	not recorded	standard_result.median_decode_tps	Run 4301ebe8 →
n	3.00 count	machine total	single	not recorded	standard_result.n	Run 4301ebe8 →

Data and citation

Download this run · OLIR JSON All numbers · CSV OLIR schema

Skilak Keep evidence, run 4301ebe8ec824fb18f5a4d1d, measured 2026-10-18T20:26:01.269182+00:00.
https://skilak.ai/keep/resources/evidence/r/4301ebe8ec824fb18f5a4d1d