A 9-billion-parameter open model fine-tuned with reinforcement learning outperformed several frontier-model configurations on a catalog-review task, according to results published by FermiSense. The company presents the experiment as evidence that a smaller specialist can beat general-purpose systems when it is trained against a scored version of a specific business workflow.
The reported test measured each system as a share of the maximum achievable catalog-review score. FermiSense says its trained model reached a level 13.5% above the best frontier configuration, which scored 76.9%, and 36% above the specialist model's own untrained baseline of 64.2%. Five frontier systems using optimized prompts finished within one tenth of a point of each other, according to the account.
Cost was the other major difference. FermiSense reports an inference price of $0.50 per 1,000 reviewed listings for the specialist, compared with $34 for the strongest frontier baseline. It says the trained system remained 40 times cheaper than the least expensive frontier option in its comparison. At a hypothetical volume of 40 million decisions per day, the article estimates annual costs of roughly $7 million for the specialist and $500 million for the strongest frontier configuration.
These are company-published results from one workflow, not an independent general benchmark. Performance depends on the task definition, scoring process, training data, model choices and deployment assumptions. The evidence supplied does not establish that the same model would outperform broad systems on unrelated work.
FermiSense describes the approach as a repeatable three-part pattern: select an open-source base model, collect proprietary examples from the real task, and use reinforcement learning against an evaluator that can score the workflow. That differs from relying entirely on prompts or retrieval to provide company context on every call. The organization instead encodes repeated task behavior into the specialist model.
The company does not argue that specialist models should replace frontier services everywhere. Its proposed workflow begins with general-purpose models to establish feasibility and generate examples. Once a process reaches volume, operators can train a cheaper specialist for the portions that depend heavily on internal data and procedures. Either model can still call the other when a task requires different capabilities.
The experiment illustrates a specific trade-off in enterprise AI. Large general systems offer wide competence and a fast path to prototyping, while a smaller model can be optimized for a narrow, measurable decision. For organizations considering that route, the reported gain makes evaluation design central: without a reliable task score, there is no sound basis for training rewards or for judging whether the specialist actually improves the production workflow.
