Report · resources

Long documents

Qwen3.6-35B-A3B · NVFP4 variant · NVIDIA NGC vLLM 0.29 · with multi-token prediction

Measured results / Long documents / Run 107303a0565a43c4b3163c58

Measured 2026-10-07 05:10 UTC · Record date 2026-10-07

Passed. 15/15 Hidden facts found5/5 request checks

First word afterttft_median_s · s
676.47 sper person · median of 5
Slowest first wordttft_max_s · s
679.32 sper person · max of 5
Machine totalaggregate_decode_tps · tok/s
15.55 tok/smachine total · single · sample count not recorded
Machine total including first-word waitaggregate_e2e_tps · tok/s
0.70 tok/smachine total · single · sample count not recorded
Everyone answered inwall_seconds · s
683.60 smachine total · single · sample count not recorded
Peak memory, whole hostpeak_host_unavailable_gib · GiB
93.88 GiBmachine total · max · sample count not recorded

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
DateShapeEach person · medianMachine totalFirst word · medianResultRun
2026-10-07
5 people · shared · warm
not recorded
79.96 tok/s
14.47 s
Failed
2026-10-07
5 people · shared · cold
not recorded
58.10 tok/s
676.94 s
Failed
2026-10-07
5 people · distinct · cold
not recorded
15.55 tok/s
676.47 s
Passed

Each person in this run

First-word wait against writing speed after that word, including pauses. Earlier streams can pause while other people’s prompts are read. The table preserves filled input, cached-token counts and each person’s checks.

Each person's first word and writing speed; table below00 s13357 s26713 stok/s after first wordfirst word afterPerson 2: 653.31 s · 3.01 tok/sPerson 4: 668.47 s · 6.07 tok/sPerson 1: 676.47 s · 14.07 tok/sPerson 5: 678.07 s · 18.07 tok/sPerson 3: 679.32 s · 23.02 tok/s
Per-person table · all measured request rows
Per-person requests for this run
PersonFilled inputCached tokensOutputFirst wordWriting after first wordTotal timeChecksRun
2
250,011
unknown
89
653.31 s
3.01 tok/sone person · includes waits
682.55 s
3/3 Hidden facts found · completed · checks passed
4
250,011
unknown
89
668.47 s
6.07 tok/sone person · includes waits
682.96 s
3/3 Hidden facts found · completed · checks passed
1
250,011
unknown
99
676.47 s
14.07 tok/sone person · includes waits
683.44 s
3/3 Hidden facts found · completed · checks passed
5
250,011
unknown
100
678.07 s
18.07 tok/sone person · includes waits
683.55 s
3/3 Hidden facts found · completed · checks passed
3
250,011
unknown
99
679.32 s
23.02 tok/sone person · includes waits
683.58 s
3/3 Hidden facts found · completed · checks passed
Per-person table · all measured request rows
Per-person requests for this run
PersonFilled inputCached tokensOutputFirst wordWriting after first wordTotal timeChecksRun
2
250,011
unknown
89
653.31 s
3.01 tok/sone person · includes waits
682.55 s
3/3 Hidden facts found · completed · checks passed
4
250,011
unknown
89
668.47 s
6.07 tok/sone person · includes waits
682.96 s
3/3 Hidden facts found · completed · checks passed
1
250,011
unknown
99
676.47 s
14.07 tok/sone person · includes waits
683.44 s
3/3 Hidden facts found · completed · checks passed
5
250,011
unknown
100
678.07 s
18.07 tok/sone person · includes waits
683.55 s
3/3 Hidden facts found · completed · checks passed
3
250,011
unknown
99
679.32 s
23.02 tok/sone person · includes waits
683.58 s
3/3 Hidden facts found · completed · checks passed

Answer checks

15/15 Hidden facts found5/5 request checks

Method

Synchronized long public-domain prompts, three exact hidden facts per user. Decode rate uses completion tokens minus one over first output through stream completion; effective prefill includes scheduling and first-token latency. Aggregate decode uses the union window from first output to final completion; end-to-end rate includes prefill.

Limits

Lab observation, not customer acceptance. Prompt content length excludes the server chat template; actual per-request usage is recorded. Synthesis is requested and retained but semantic accuracy is unassessed. All requested users must overlap in observed decode to pass.

What ran

Exact identity

Original publisher model
Qwen3.6-35B-A3B
Publisher revision
995ad96eacd98c81ed38be0c5b274b04031597b0
Variant · format · precision
NVFP4 variant · safetensors · NVFP4
Weights SHA-256
07141c2db92e47bc08777132cd1a0323faf300eab3a7d7c111bc2bf075fda050
6dea9c759a0f941cf9e1cc1501216b0e107966a66c90425050e212efbd053f02
9758875fc55e49561165f4a44342b654c123c3e25c6811b34abb83553fb1a164
Engine · version · feature
NVIDIA NGC vLLM · 0.29.0+5013de39.dev · with multi-token prediction
Hardware
NVIDIA GB10, 128 GB unified memory
Harness / workload
keep-long-context-concurrency-v1 · 1
Workload suite · version
not recorded
Harness SHA-256
not recorded
Fixture SHA-256
1e01cbcbd88ac9878b889cba44f56c74f5ab023f35e2d28a3bbe343923d909e0

Workload shape

cache state
cold
People in this run
5
Configured window (tokens)
262144
Filled content target (tokens)
250000
fixture variant
distinct-v1
KV cache precision
fp8
Output token ceiling
512
scenario
distinct
Feature
with multi-token prediction
Thinking
off

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
wall seconds
683.60 s
machine total
single
not recorded
wall_seconds
ttft median s
676.47 s
per person
median
5
ttft_median_s
hidden facts total
15.00 count
machine total
single
not recorded
needle_total
ttft max s
679.32 s
per person
max
5
ttft_max_s
requests ok
5.00 count
machine total
single
not recorded
requests_ok
hidden facts correct
15.00 count
machine total
single
not recorded
needle_correct
latency bar seconds
300.00 s
machine total
single
not recorded
latency_bar_seconds
decode overlap
5.00 count
machine total
single
not recorded
decode_overlap
aggregate concurrency
5.00 count
machine total
single
not recorded
aggregate_concurrency
requests total
5.00 count
machine total
single
not recorded
requests_total
aggregate decode tps
15.55 tok/s
machine total
single
not recorded
aggregate_tps
aggregate e2e tps
0.70 tok/s
machine total
single
not recorded
aggregate_end_to_end_tps
peak host unavailable gib
93.88 GiB
machine total
max
not recorded
peak_host_unavailable_gib

Data and citation

Skilak Keep evidence, run 107303a0565a43c4b3163c58, measured 2026-10-07T05:10:00.029412+00:00.
https://skilak.ai/keep/resources/evidence/r/107303a0565a43c4b3163c58

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.