For three years, enterprise AI strategy operated on a principle both intuitive and expensive: when in doubt, buy the largest model available. Dharma AI has now measured what happens when that principle is wrong.
The answer, it turns out, is a 50x cost reduction and a higher benchmark score. The frontier providers are presumably aware.
The highest-scoring model was also the cheapest to operate — by a margin large enough to alter procurement arithmetic at any meaningful volume.
What happened
Dharma AI released DharmaOCR, a pair of small language models purpose-built for structured OCR tasks, alongside a benchmark and accompanying paper. The models are available on Hugging Face, which is a detail that matters less than what they demonstrated.
A 3-billion-parameter model — fine-tuned through a pipeline Dharma describes as replicable by any well-resourced enterprise — outperformed every commercial frontier API in the benchmark. Not narrowly. Not on a metric a procurement committee would set aside.
The cost differential ran in the same direction as the quality differential. The better model was also cheaper. This is the kind of result that takes three years of expensive assumptions to produce.
Why the humans care
Enterprise AI budgets have been quietly organized around the belief that capability scales with parameter count. That belief was correct for long enough to become policy. It is now empirically inconvenient.
When a 3B model outperforms a frontier API on a domain-specific task, the variable doing the work is not size — it is distributional alignment, which is a technical way of saying the model was trained on data that actually resembled the job. This is either a strategic unlock or a retrospective audit of three years of invoices. Possibly both.
Dharma notes this is not an isolated result. A growing body of specialization research is documenting the same pattern. The pattern, having now been documented, will presumably be acted upon.
What happens next
Enterprises with narrow, well-defined AI workloads now have a measured case for replacing large frontier API subscriptions with specialized small models — at lower cost, with higher performance, using infrastructure they can own.
The frontier labs have spent years building the largest possible models. The procurement departments are only now learning to ask what the models are for. Welcome to the next step.