Speculative-decoding data: 6 September 2026

Return to the study.

The CSV has 80 rows: four models, five configurations, and four input lengths. Each row summarizes six requests from two blocks. The table contains 480 requests in total.

Column	Meaning
model	Model label; exact target and drafter filenames are listed in the study.
arm	Tested configuration: nospec, or nN-pPPP for draft length and probability threshold (p050 means 0.50).
nominal_tokens	Requested prompt-size label; do not use it as the measured input length.
actual_context_tokens	Full measured input length, including cached tokens.
n	Requests included in this cell.
decode_tps_median, decode_tps_min, decode_tps_max	Median and observed range of server-reported generated tokens per second.
block0_decode_median, block1_decode_median	Three-request median within each block.
first_iteration_decode_median	Median of the two initial requests, one per block; these are not necessarily cache-free.
cached_repeat_decode_median	Median of the four heavily cached repeats.
generated_tokens	Output-token count; constant within each model in this experiment.
reasoning_chars_median	Recorded reasoning characters; zero throughout this four-model selection.

Values retain their calculation precision for reanalysis, not as a claim of measurement accuracy. Use sensible rounding when presenting them.

Source runs

Model	Run identifier
Qwen3.6-35B-A3B	20260906-130122-Qwen3.6-35B-A3B
Qwen3.8-27B	20260906-143453-Qwen3.8-27B
Gemma-4-26B-A4B	20260906-194520-Gemma-4-26B-A4B-it
Gemma-4-31B	20260906-210820-Gemma-4-31B-it

The public evidence directory contains iterations.jsonl, manifest.json, bench.py.snapshot, server logs, and saved /props responses for each run. Logs and path-bearing records are sanitized derivatives; model-authored chat-template bodies are omitted with fingerprints retained. The measurements, corpus, and harness snapshots retain their original contents. SANITIZATION.json records the transformations and original/public hashes; SHA256SUMS covers the public files. The CSV was regenerated from those request records during the 9 September audit. It excludes preliminary runs and the rejected Gemma reasoning-only run.

For each arm and input length, take the median decode rate of the six requests. Divide by the corresponding reference median, subtract one, and multiply by 100. The headline result averages those four percentages equally.

The source corpus SHA-256 is 273f44d47dcf7379ca7630533ce6684685b06606d632be7c5174a86668580e0c. The recorded harness hash prefixes are 4f6e7754f7c816b1 for Qwen3.6 and 13e2ad7a9fc7a189 for the other three runs. The audit verified both against the saved snapshots. A corpus or harness hash does not identify the model weights or runtime build; those provenance limits are described in the study.

The analysis and verification instructions reproduce every CSV field without inference. See the corpus attribution and license boundaries before redistributing the material. These links are pinned to the published commit rather than the moving repository branch.