Report · resources

Rolling chat

Qwen3.8-27B · INT4 variant · NVIDIA NGC vLLM 0.29

Measured results / Rolling chat / Run 8a22159d4ed2420f8fe7735f

Measured 2026-10-07 21:47 UTC · Record date 2026-10-07

Failed. 50/50 turns

Answer checks

50/50 turns

Rolling-chat turn checks
All rolling-chat turns
Person · turnFilled inputFirst wordRecent · pinned · oldResultRun
1 · 1
20208
21.35 s
true · true · recalled
Passed
1 · 2
40334
26.00 s
true · true · recalled
Passed
1 · 3
60462
48.95 s
true · true · recalled
Passed
1 · 4
80595
56.14 s
true · true · recalled
Passed
1 · 5
100723
82.31 s
true · true · recalled
Passed
1 · 6
120853
91.19 s
true · true · recalled
Passed
1 · 7
80765
212.87 s
true · true · recalled
Passed
1 · 8
100891
36.44 s
true · true · recalled
Passed
1 · 9
121021
70.18 s
true · true · recalled
Passed
1 · 10
80771
174.01 s
true · true · recalled
Passed
2 · 1
20214
48.61 s
true · true · recalled
Passed
2 · 2
40338
55.96 s
true · true · recalled
Passed
2 · 3
60465
82.20 s
true · true · recalled
Passed
2 · 4
80597
93.23 s
true · true · recalled
Passed
2 · 5
100728
122.79 s
true · true · recalled
Passed
2 · 6
120852
118.01 s
true · true · recalled
Passed
2 · 7
80772
235.59 s
true · true · recalled
Passed
2 · 8
100900
65.22 s
true · true · recalled
Passed
2 · 9
121034
70.11 s
true · true · recalled
Passed
2 · 10
80777
269.82 s
true · true · recalled
Passed
3 · 1
20209
81.10 s
true · true · recalled
Passed
3 · 2
40339
93.25 s
true · true · recalled
Passed
3 · 3
60471
127.11 s
true · true · recalled
Passed
3 · 4
80597
119.55 s
true · true · recalled
Passed
3 · 5
100730
115.67 s
true · true · recalled
Passed
3 · 6
120860
109.18 s
true · true · recalled
Passed
3 · 7
80775
206.20 s
true · true · recalled
Passed
3 · 8
100903
94.37 s
true · true · recalled
Passed
3 · 9
121033
63.80 s
true · true · recalled
Passed
3 · 10
80770
177.92 s
true · true · recalled
Passed
4 · 1
20212
23.82 s
true · true · recalled
Passed
4 · 2
40341
126.19 s
true · true · recalled
Passed
4 · 3
60478
118.53 s
true · true · recalled
Passed
4 · 4
80606
114.78 s
true · true · recalled
Passed
4 · 5
100735
107.12 s
true · true · recalled
Passed
4 · 6
120860
89.24 s
true · true · recalled
Passed
4 · 7
80767
275.18 s
true · true · recalled
Passed
4 · 8
100904
67.31 s
true · true · recalled
Passed
4 · 9
121033
64.34 s
true · true · recalled
Passed
4 · 10
80782
143.37 s
true · true · recalled
Passed
5 · 1
20214
108.41 s
true · true · recalled
Passed
5 · 2
40342
119.35 s
true · true · recalled
Passed
5 · 3
60472
113.27 s
true · true · recalled
Passed
5 · 4
80602
109.21 s
true · true · recalled
Passed
5 · 5
100732
90.44 s
true · true · recalled
Passed
5 · 6
120864
79.00 s
true · true · recalled
Passed
5 · 7
80777
178.91 s
true · true · recalled
Passed
5 · 8
100907
66.35 s
true · true · recalled
Passed
5 · 9
121037
64.14 s
true · true · recalled
Passed
5 · 10
80778
105.55 s
true · true · recalled
Passed

Method

Staggered multi-turn fictional archive, exact engine-tokenized rolling gateway; independent recent, pinned and old probes each turn.

Limits

Bounded tuning pilot; incomplete window growth cannot pass the full workload grade. Lab retrieval and rolling-history evidence only. Bounded literal user-source excerpts, with verified source hashes; details outside them remain unavailable. Shared engine cache intervals include overlapping users and are not per-user hits. Gateway path, not authenticated document ingestion or customer acceptance.

What did not work

  • Bounded tuning pilot; incomplete window growth cannot pass the full workload grade.

What ran

Exact identity

Original publisher model
Qwen3.8-27B
Publisher revision
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
Variant · format · precision
INT4 variant · safetensors · INT4
Weights SHA-256
not recorded
Engine · version · feature
NVIDIA NGC vLLM · 0.29.0+5013de39.dev · plain
Hardware
NVIDIA GB10, 128 GB unified memory
Harness / workload
keep-rolling-context-v1 · 1
Workload suite · version
not recorded
Harness SHA-256
not recorded
Fixture SHA-256
not recorded

Workload shape

People in this run
5
Configured window (tokens)
262144
Configured window (tokens)
262144
Output token ceiling
256

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
roll ttft max seconds
275.18 s
per person
max
not recorded
roll_ttft_max_seconds
requests ok
50.00 count
machine total
single
not recorded
requests_ok
roll ttft p50 seconds
192.55 s
per person
median
not recorded
roll_ttft_p50_seconds
rolled off abstained
0.00 count
machine total
single
not recorded
rolled_off_abstained
rolled off invalid
0.00 count
machine total
single
not recorded
rolled_off_invalid
rolled off wrong
0.00 count
machine total
single
not recorded
rolled_off_wrong
normal ttft max seconds
127.11 s
per person
max
not recorded
normal_ttft_max_seconds
rolled off correct
20.00 count
machine total
single
not recorded
rolled_off_correct
normal ttft p50 seconds
85.78 s
per person
median
not recorded
normal_ttft_p50_seconds
rolled off total
20.00 count
machine total
single
not recorded
rolled_off_total
requests total
50.00 count
machine total
single
not recorded
requests_total

Data and citation

Skilak Keep evidence, run 8a22159d4ed2420f8fe7735f, measured 2026-10-07T21:47:51.419664+00:00.
https://skilak.ai/keep/resources/evidence/r/8a22159d4ed2420f8fe7735f

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.