Narrow distilled small models will reach at least 10 named company deployments — resolves 2027-07-14

The debate: research already shows that small models trained by a larger one, known as distillation, and focused on one narrow task can beat far bigger models on problems such as math proofs, but the question is whether companies will actually put them into production.

Why it matters: if narrow small models work commercially, much corporate work no longer has to pay large-model prices, and the cost structure of AI inference, the computing that produces a model's answers, shifts with it.

Who is on each side: one camp sees this as a lab result while companies keep using general large models, and SecondSource expects named commercial cases of narrow distilled models to add up, though with only 40% confidence.

SecondSource’s reading: by 2027-07-14, publicly verifiable commercial deployments of narrow-domain distilled small models will total at least 10. Each case needs a named company and a named use; vague claims that some big firm uses one do not count. We put our confidence at 40%.

The research backing: the DeepSeek-R1 paper says the reasoning patterns that emerge in large models can be systematically distilled into small ones, and QED-Nano, a 4-billion-parameter open model, beats GPT-OSS-120B on olympiad-level theorem proving at a small fraction of the inference cost.

We deliberately use an absolute threshold of 10 rather than a growth rate, because a growth rate needs a count of today's cases, and we have not verified that number.

arXiv (DeepSeek-AI, collective lab authorship)
“we show that the reasoning abilities of LLMs can…”

arXiv (LM-Provers / Qu / Setlur / Tunstall / Aviral Kumar)
“surpasses the proof-generation performance of much…”

arXiv
“both pure and hybrid Mamba models can scale their…”

Compile cases from mainstream tech press coverage and lab and cloud vendor case study pages, listing each named narrow-domain distilled small model deployment individually and counting them.

What would overturn it

A cumulative count below 10 weakens the reading that narrow-domain distillation is spreading commercially.