Much of the AI conversation today revolves around scale. More parameters. Larger context windows. Bigger clusters. Longer-running reasoning chains.
But one of the most interesting lessons we have learned recently points in the opposite direction: sometimes smaller models win.
Our teams built a proprietary 1-billion-parameter foundation model in our own GPU environment. Rather than starting with an external model, we pre-trained a foundation model on proprietary transaction data, creating a reusable AI foundation tailored to the patterns, relationships, and structures embedded within those transactions.
We then adapted that foundation model to fraud detection. Rather than reaching for the largest general-purpose model available, we explored what happens when a model is purpose-built for a specific domain and optimized for a clearly defined business objective.
That distinction matters. Pre-training creates a reusable foundation that can support multiple downstream applications, while adaptation allows the model to specialize for a particular task. In our case, a 1-billion-parameter model proved capable of operating at enterprise scale while remaining efficient enough for production deployment.
The results have been encouraging. The model identifies more fraud, improves detection accuracy by 24% compared with previous approaches, and reduces false alerts by 14%, all while operating within a framework of human oversight and strong governance controls.
Three Engineering Lessons
1. Domain-Specific Knowledge Can Outweigh Scale
A model built for a specialized domain does not need to know everything. By pre-training on proprietary transaction data, the model develops domain-specific representations that capture behavioral patterns, transactional relationships, risk signals, and operational characteristics that matter most to the task at hand.
For many enterprise use cases, domain relevance can be more valuable than additional parameters.
2. Smaller Models Have Real Operational Advantages
Model quality is not the only metric that matters in production. Latency, throughput, infrastructure cost, and operational complexity are equally important considerations.
Smaller models are typically faster to serve, less expensive to run, and easier to scale. In fraud detection, where decisions often need to be made while a transaction is being processed, those advantages translate directly into business value.
3. Pre-Training Creates a Reusable Asset
What excites us most is not a single fraud use case.
By pre-training a foundation model on proprietary transaction data, we created a reusable foundation that can be adapted to additional applications. The value is not limited to fraud detection. The same foundation can support multiple downstream tasks, allowing organizations to leverage pre-training investments across a broader set of business problems.
The broader lesson is that AI progress is not solely about building ever-larger models. In many enterprise environments, the winning strategy may be to pre-train the right-sized foundation model on proprietary data and then adapt it to high-value business problems.
The future may not be exclusively about scaling up. In many cases, it may be about scaling intelligently.