Report · resources

Team load

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

Measured results / Team load / Run ed509c2d71d243a1b4d18c7e

Measured 2026-10-07 04:34 UTC · Record date 2026-10-07

Passed. 3/3 Hidden facts found1/1 probes1/1 request checks

Writing speed, each persondecode_tps · tok/s
90.70 tok/sper person ·
Writing speed, one personsingle_user_decode_tps · tok/s
90.70 tok/sper person · single · sample count not recorded
First word afterttft_median_s · s
10.40 sper person ·
Slowest first wordttft_max_s · s
10.40 sper person ·
Machine totalaggregate_decode_tps · tok/s
90.70 tok/smachine total · single · sample count not recorded
Machine total including first-word waitaggregate_e2e_tps · tok/s
31.93 tok/smachine total · single · sample count not recorded
Everyone answered inwall_seconds · s
16.04 smachine total · single · sample count not recorded
Peak memory, whole hostpeak_host_unavailable_gib · GiB
91.99 GiBmachine total · max · sample count not recorded
mtp acceptance ratemtp_acceptance_rate · ratio
0.83 ratiomachine total · single · 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
20 people
6.98 tok/s
76.29 tok/s
115.25 s
Passed
2026-10-07
16 people
8.27 tok/s
83.08 tok/s
92.15 s
Passed
2026-10-07
8 people
21.05 tok/s
149.84 tok/s
54.98 s
Passed
2026-10-07
1 people
90.70 tok/s
90.70 tok/s
10.40 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 s525 s10411 stok/s after first wordfirst word afterPerson 1: 10.40 s · 90.70 tok/s
Per-person table · all measured request rows
Per-person requests for this run
PersonFilled inputCached tokensOutputFirst wordWriting after first wordTotal timeChecksRun
1
40,103
unknown
512
10.40 s
90.70 tok/sone person · includes waits
16.04 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
1
40,103
unknown
512
10.40 s
90.70 tok/sone person · includes waits
16.04 s
3/3 Hidden facts found · completed · checks passed

Answer checks

3/3 Hidden facts found1/1 probes1/1 request checks

Method

team-load-v1; distinct generated 40000-token coding contexts plus fresh nonce prefixes; synchronized streams; temperature zero; thinking off; min_tokens=max_tokens=512; separate cold hidden facts probes; no retries.

Limits

Finite lab cohort; configured 262144 context differs from measured inputs. Aggregate uses first-output to final-completion window. Whole-host memory sampled during speed only; missing counters remain unassessed. Broader coding quality and delivery acceptance unassessed.

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-team-load-v1 · 1
Workload suite · version
not recorded
Harness SHA-256
not recorded
Fixture SHA-256
e2097959b1ca0142f2b93edc3ec411e474c0a6622800dd247af7a71bd9c39e83

Workload shape

People in this run
1
Configured window (tokens)
262144
Configured window (tokens)
262144
KV cache precision
fp8
Output token ceiling
512
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
16.04 s
machine total
single
not recorded
wall_seconds
ttft median s
10.40 s
per person
median
1
ttft_median_s
hidden facts total
3.00 count
machine total
single
not recorded
needle_total
ttft max s
10.40 s
per person
max
1
ttft_max_s
requests ok
1.00 count
machine total
single
not recorded
requests_ok
hidden facts correct
3.00 count
machine total
single
not recorded
needle_correct
preemptions
0.00 count
machine total
single
not recorded
preemptions
decode overlap
1.00 count
machine total
single
not recorded
decode_overlap
aggregate concurrency
1.00 count
machine total
single
not recorded
aggregate_concurrency
requests total
1.00 count
machine total
single
not recorded
requests_total
mtp acceptance rate
0.83 ratio
machine total
single
not recorded
mtp_acceptance_rate
correctness probes passed
1.00 count
machine total
single
not recorded
correctness_probes_passed
aggregate decode tps
90.70 tok/s
machine total
single
not recorded
aggregate_tps
aggregate e2e tps
31.93 tok/s
machine total
single
not recorded
aggregate_end_to_end_tps
correctness probes total
1.00 count
machine total
single
not recorded
correctness_probes_total
single user decode tps
90.70 tok/s
per person
single
not recorded
single_user_decode_tps
decode tps
90.70 tok/s
per person
mean
1
per_request_decode_tps
forced output tokens
512.00 tokens
machine total
single
not recorded
forced_output_tokens
peak host unavailable gib
91.99 GiB
machine total
max
not recorded
peak_host_unavailable_gib
decode tps
90.70 tok/s
per person
median
1
per_person_decode_median_tps

Data and citation

Skilak Keep evidence, run ed509c2d71d243a1b4d18c7e, measured 2026-10-07T04:34:31.551279+00:00.
https://skilak.ai/keep/resources/evidence/r/ed509c2d71d243a1b4d18c7e

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.