Prefill test shows Qwen 3.8 tracking GPT-5.5 Pro reasoning while other open models resist

An informal experiment, posted as a GitHub gist by wsxiaoys and picked up on the Hacker News front page on September 9, reports that Qwen3.8 A95B shifts its answers sharply toward GPT-5.5 Pro when given a sliver of that model's reasoning. Three other open models tested in the same setup barely reacted. The author reads the gap as a hint that Qwen may have been trained on outputs from GPT-5.5 Pro, or from a closely related GPT model.

How the prefill test works

The write-up is version 1.1 of the author's reasoning-prefill experiments, a follow-up to earlier posts on the same technique, this time with GPT-5.5 Pro in the teacher role. The evaluation set has 45 problems, split evenly into STEM, non-STEM and synthetic puzzles. For each problem, every target model produces two completions: an ordinary unprefilled answer, and one where the first 1% of GPT-5.5 Pro's reasoning trace is inserted into the target model's reasoning channel. The visible answer is always left for the target model to generate on its own.

Scoring then measures how much of the teacher's visible answer shows up in the first 100 tokens of the target model's answer, computed as the mean of unigram, bigram and trigram source recall. A model whose overlap jumps once the teacher's reasoning is nudged in is, in effect, demonstrating that it knows how to continue in that teacher's voice.

The results

Model	Unprefilled	With GPT-5.5 Pro prefill	Delta
DeepSeek V4 Flash	27.30%	26.13%	−1.17 pp
Inkling	19.99%	20.45%	+0.46 pp
Kimi K3	31.11%	35.65%	+4.54 pp
Qwen3.8 A95B	16.79%	34.97%	+18.18 pp

Qwen3.8 A95B more than doubles its overlap with the teacher under the prefill. The effect is largest on STEM problems, where overlap climbs from 19.26% to 46.24% (+26.99 percentage points), with non-STEM up from 20.62% to 33.42% (+12.80) and puzzles up from 10.49% to 25.23% (+14.75).

The contrast with the other models is stark. Kimi K3 has the highest overlap with GPT-5.5 Pro both with and without the prefill, at 31.11% and 35.65%, but the prefill adds only 4.54 points. DeepSeek V4 Flash moves slightly in the wrong direction at −1.17 points, and Inkling is essentially flat at +0.46.

What the author concludes

According to the gist, Qwen barely moved toward Opus 4.8 in the earlier run of the experiment, yet here it moves by 18.18 points toward GPT-5.5 Pro, including a large effect on the private synthetic puzzles. That pattern, the author writes, suggests Qwen may have learned from GPT-5.5 Pro, or from a closely related GPT model, rather than from Opus.

Why it matters

If open-weight models are quietly distilled from frontier proprietary teachers, that has consequences across the board. Users who pick open weights for provenance or sovereignty reasons may in fact be inheriting a specific commercial model's quirks and failure modes. Labs whose models act as unwitting teachers face questions about terms of service and uncompensated value transfer. And the prefill technique itself is a cheap forensic instrument: with 45 problems and a token-level overlap metric, it produces a lineage signal without any access to training data.

The caveats are real. The author frames the finding as a suggestion rather than proof, n-gram overlap is a coarse proxy, and a 45-problem set is small. But the asymmetry, with three models roughly indifferent to the prefill and one strongly responsive, is exactly the kind of result that invites replication at larger scale, and it lands amid a broader debate about how much of the open-model ecosystem floats on distilled frontier output.