Report · resources

Tuning sweep

Qwen3.8-27B · NVFP4 variant · SGLang · speculative decoding

Measured results / Tuning sweep / Run f8a9c1e519df761b52177b08

Measured 2026-10-05 15:43 UTC · Record date 2026-10-05

Partial.

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
2026-10-05
2 people
not recorded
59.16 tok/s
not recorded
Partial
2026-10-05
unknown people
not recorded
not recorded
not recorded
Partial
2026-10-05
unknown people
not recorded
not recorded
not recorded
Partial

Answer checks

No answer-quality suite was graded for this run.

Method

Sixteen exact-output tasks and repeated greedy controls per candidate; warm four-request code/chat throughput includes first-token time and final drain.

Limits

Earlier attempt; comparison or runtime-boundary checks stopped the sweep. No winner promoted. Every attempted and unrun candidate is retained.

What did not work

  • Earlier attempt; comparison or runtime-boundary checks stopped the sweep.

What ran

Exact identity

Original publisher model
Qwen3.8-27B
Publisher revision
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
Variant · format · precision
NVFP4 variant · safetensors · NVFP4
Weights SHA-256
not recorded
Engine · version · feature
SGLang · version not recorded · speculative decoding
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

Configured window (tokens)
262144
Feature
speculative decoding

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
tuning candidates
9.00 count
machine total
single
not recorded
tuning_candidates
rejected candidates
6.00 count
machine total
single
not recorded
rejected_candidates

Data and citation

Skilak Keep evidence, run f8a9c1e519df761b52177b08, measured 2026-10-05T15:43:15.587639+00:00.
https://skilak.ai/keep/resources/evidence/r/f8a9c1e519df761b52177b08

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.