← Glossary Company
Fermisense
A consultancy that published a widely read 2026 result showing a fine-tuned 9 billion parameter open model beating frontier models on one narrow business task.
Fermisense trained a small open model on e-commerce catalogue review using reinforcement learning, and reported 87.3 percent accuracy against 76.9 for the best frontier setup it tested. The write-up did well on Hacker News because it argued something people wanted to hear: you do not always need the biggest model.
The caveat is in the word narrow. The result holds on one well-defined, repetitive task with plenty of examples to learn from, which is exactly the situation where a small specialised model wins. It is not evidence that small models beat frontier models in general.
Mentioned in