Report · resources

Long documents

Qwen3.8-27B · Mirai S trellis with Q8_0 MTP GGUF variant · Mirai S CUDA dev · with multi-token prediction

Measured results / Long documents / Run 6a38761f11ed4674a966f55c

Measured 2026-10-09 15:28 UTC · Record date 2026-10-09

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

First word afterttft_median_s · s
1,270.43 sper person ·
Slowest first wordttft_max_s · s
1,270.43 sper person ·
Peak memory, whole hostpeak_host_unavailable_gib · GiB
12.85 GiBmachine total · max · sample count not recorded

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 s2667 s31,334 stok/s after first wordfirst word afterPerson 1: 1,270.43 s · 2.71 tok/s
Per-person table · all measured request rows
Per-person requests for this run
PersonFilled inputCached tokensOutputFirst wordWriting after first wordTotal timeChecksRun
1
250,011
unknown
112
1,270.43 s
2.71 tok/sone person · includes waits
1,311.44 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
250,011
unknown
112
1,270.43 s
2.71 tok/sone person · includes waits
1,311.44 s
3/3 Hidden facts found · completed · checks passed

Answer checks

3/3 Hidden facts found1/1 request checks

Method

Frozen three-hidden facts fixture; native token count verified; owned managed container recreated cold. One request, no retries.

Limits

Retrieval completion and the 300-second first-token bar are independent. Synthesis semantics, broader quality and customer acceptance are unassessed.

What ran

Exact identity

Original publisher model
Qwen3.8-27B
Publisher revision
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
Variant · format · precision
Mirai S trellis with Q8_0 MTP GGUF variant · GGUF · Mirai S trellis with Q8_0 MTP
Weights SHA-256
af3e2d508cb12ebf68688a0756da6bcb4cdfceca3883f5751666d5e4fb34a334
Engine · version · feature
Mirai S CUDA · 0.2.0-dev-vector-disabled · with multi-token prediction
Hardware
NVIDIA GeForce RTX 3060, 12 GB VRAM
Harness / workload
not recorded
Workload suite · version
not recorded
Harness SHA-256
not recorded
Fixture SHA-256
not recorded

Workload shape

cache state
cold
People in this run
1
Configured window (tokens)
262144
Filled content target (tokens)
250000
fixture variant
original-tokenizer-needles-v1
KV cache precision
q8_0
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
ttft median s
1,270.43 s
per person
median
1
ttft_median_s
hidden facts total
3.00 count
machine total
single
not recorded
needle_total
ttft max s
1,270.43 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
latency bar seconds
300.00 s
machine total
single
not recorded
latency_bar_seconds
decode overlap
1.00 count
machine total
single
not recorded
decode_overlap
requests total
1.00 count
machine total
single
not recorded
requests_total
peak host unavailable gib
12.85 GiB
machine total
max
not recorded
peak_host_unavailable_gib

Data and citation

Skilak Keep evidence, run 6a38761f11ed4674a966f55c, measured 2026-10-09T15:28:53.860792+00:00.
https://skilak.ai/keep/resources/evidence/r/6a38761f11ed4674a966f55c

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.