Smaller AI models can handle legal documents as well as expensive ones

Researchers tested whether smaller AI models (under 10 billion parameters) can analyze legal contracts and court cases as effectively as much larger, more expensive models like GPT-4o-mini. They found that a smaller model using a specific architecture can match the larger model's performance while costing $62 total to evaluate, suggesting that legal AI deployment doesn't require the massive models that are expensive and slow to run.

If smaller models genuinely perform as well as large ones on legal tasks, law firms and legal departments could build AI systems that are faster, cheaper to operate, and don't require sending sensitive contracts to third-party cloud providers — removing a major practical barrier to actual legal AI deployment rather than just research demos.