Joakimpalm-Zen/Qwen3-speculative-pair-report

Qwen3 speculative pair report: 0.6B draft + 8B target, measured acceptance

<!-- BEGIN xyntetik-dataset-summary -->

Research evidence dataset. No model weights. Part of the collection Xyntetik Research: Runner Compatibility Reports on this account, produced with Xyntetik Runner.

Dataset summary

Question tested. What the acceptance rate of a Qwen3-0.6B draft against a Qwen3-8B target actually is across draft depths, whether the engine's printed tokens-per-round figure can be tuned on (it cannot), and whether the 8B target's own quants clear the house bar.

Models involved.

• —Qwen/Qwen3-8B-GGUF Q80 (target and reference), Q6K, Q5KM, Q4KM
• —Qwen/Qwen3-0.6B GGUF exports at q4_0 and bf16 as drafts

Method. Whole-model fidelity of the target's quants at 400 teacher-forced positions; then a draft-k sweep (2, 4, 8) on the CPU path at temperature 0, 96 tokens, same prompt, counting drafted and accepted tokens per round.

What each file contains.

• —README.md: the report and both tables
• —gates/qwen3-8b-quant-fidelity-report.json: the quant-fidelity instrument's machine record for the Qwen3-8B ladder (schema xyntetik.quant-fidelity.v1, runner 0.1.15-alpha, 2026-08-12); paths inside are the measurement box's and are left as recorded
• —gates/spec-pair-sweep.txt: the raw sweep lines, one per (draft, k): rounds, drafted, accepted, printed tok/round

Reproduction. Runner's --draft / --draft-k on the CPU path with the files named by SHA-256; the target ladder with scripts/quant-fidelity.py.

Result. Best configuration is draft-k 2 at 44.1% acceptance; printed tok/round rises with k while acceptance collapses to 14.8%, so tok/round is not an acceptance rate. The q4_0 draft beats the bf16 draft at every k.

Links.

• —Runner: https://github.com/Joakimpalm-Zen/xyntetik-runner
• —Former Model-typed location of this report: https://huggingface.co/Joakimpalm-Zen/Qwen3-speculative-pair-report <!-- END xyntetik-dataset-summary -->

The report below is the original card of `Joakimpalm-Zen/Qwen3-speculative-pair-report` (Model repository, retired as a Model on 2026-09-15 because it holds no weights); its text is unchanged.

This repository contains no model weights. It is a measurement report over Qwen's first-party files; download them from Qwen/Qwen3-8B-GGUF and Qwen/Qwen3-0.6B (GGUF exports). The measured target files, verified by SHA256 prefix before measurement:

The target's own fidelity, so the pair stands on measured ground

Whole-model fidelity of Qwen3-8B's quants against their own Q8_0 (400 teacher-forced positions; bar: margin-qualified top-1 >= 97% AND mean KLD <= 0.05; zero point exact):

One behavioural observation from the tool-fidelity instrument, reported with its caveat: on a 16-request agent matrix at temperature 0, Qwen3-8B produces the demanded structured output in 3 of 14 cases at its own Q8_0 reference and calls an offered tool in only 2, preferring to answer in prose. The runs are byte-for-byte deterministic; this is model behaviour on that matrix, not noise. Whether declining an offered tool should count against schema conformance at all is an open scoring question, so the number is reported as an observation, not a verdict.

Acceptance, measured the way that is actually informative

CPU path (--draft is silently ignored at full GPU offload), Qwen3-8B-Q8_0 target, temp 0, 96 tokens, same prompt throughout.

Read the two right-hand columns against each other. The printed tok/round rises monotonically with k, from 1.88 to 2.18, so a reader tuning on it picks k=8. The actual acceptance rate collapses over the same sweep, from 44.1% to 14.8%. At k=8 the draft model burns 3.5x the drafting compute to move the printed figure by 0.05 tokens per round. tok/round is not the acceptance rate and must never be quoted as one; both traps are documented in qwen3-30b-a3b-expert-prune-results-2026-08-11.md.

Best configuration measured: --draft-k 2, 44.1% acceptance.

One measured curiosity: the q4_0 draft beats the bf16 draft at every k (44.1 vs 37.3 at k=2). A more precise draft is not a better draft here. What matters is whether the small model's opinion coincides with the large one's, and extra draft precision does not buy that. If it holds up on other pairs it is a useful cost result, because it means the draft can be the cheap quant.