A 0.6B model running on-device can match much of what an 8B reasoning model produces — but only for some tasks, and with a real catch.
Researchers distilled a DeepSeek-R1 8B reasoning model into a Qwen3-0.6B student using QLoRA across three seeds, then benchmarked the result against several baselines: few-shot prompting, constrained decoding, a same-size non-reasoning teacher, and a larger managed pipeline. The task was structured news enrichment — map each article to a JSON object with a short summary and five categorical labels. The student model ran each article in roughly 0.8 seconds versus the teacher's 39 seconds, recovered 58% of the quality gap between the untuned base and the full teacher, and beat constrained decoding by 16.8 points on summary quality.
The interesting finding is not the speed gain — it is what does and does not transfer. A same-size instruction-tuned teacher produced a student no better than the untuned base, which means the writing quality improvement traces specifically to the reasoning nature of the teacher, not its parameter count. Label diversity, meanwhile, transferred better from a larger managed pipeline than from the reasoning model. No single teacher wins every field.
There is a grounding problem worth flagging. On the 22 short, thin-source articles in the 93-item test set, the reasoning-lineage student fabricated more often than the instruction-tuned one — 55 faithful outputs versus 74. The researchers describe this as a directional finding on a small subgroup, not a statistically significant aggregate result, which is honest but also the kind of caveat that tends to get dropped when someone ports this into production. The upshot is a routing map: choose your distillation source by field, not by overall benchmark score.