Small models on proprietary corpora: a 9B model at 98% lower cost scores near the frontier

HR and payroll platform Deel released two in-house models on 21 October. One is a 9B open-weights model that scored 70 on the company’s own 601-question benchmark, against 74 for the frontier model Claude Opus 5.5 — but at a median response of 7.7 seconds versus 14.8 to 28.4 seconds, and a claimed cost about 98% lower.

A small score gap, a large bill gap

The value of this case is not the slogan of “small but strong”; it is the comparison: the same business benchmark, the same question set, a four-point difference in score, an order-of-magnitude difference in cost and latency. For high-frequency, low-risk question answering, those four points are often not worth ten times the price.

Where the corpus comes from

K1 was adapted on about 1.2 billion tokens of material drawn from more than 2,000 Deel internal specialists; P1 is the payroll model, built on Nemotron and trained by AI21 Labs. In other words, the moat is business corpus that others cannot obtain, not model architecture. That also explains why the same open-source base produces very different results at different companies.

When it still will not do

For long-horizon agents, high-stakes decisions and regulated settings, the gap between frontier and mid-tier models is real. The sensible approach is layered: small models cover the surface — high volume, high tolerance, fast and cheap; frontier models cover the points — low volume, expensive to get wrong.

What it means for enterprises

• First take stock of how much usable, high-quality corpus you hold; that decides whether this path is open to you at all.
• Layer the tasks; do not use one model for every scenario.
• Building your own is not the same as training your own: fine-tuning plus private deployment already covers most internal scenarios.
• Work out cost per task, not per token — latency and rework are costs too.