Small Models, Big Results: Training a Finance Agent for Under $500 — Charles Dickens, Snorkel AI
A 4B parameter model outperformed a 235B model on finance tasks, trained for under $500.
“reliability and specialization may outweigh the sheer number of parameters”
Snorkel AI and UC Berkeley's Sky Computing Lab fine-tuned a 4B parameter Qwen model on high-quality synthetic financial data, achieving ~60% task success versus 51% for a 235B parameter model on the FinQA benchmark — at a training cost under $500. The result challenges the prevailing assumption that scale is the primary driver of enterprise AI performance. It reinforces a growing consensus that domain-specific data quality can unlock disproportionate gains from small, deployable models.