Samaya AI has released Samaya Sterling, a post-trained 3B-active finance research agent model that outperforms Claude Fable 5.1 on the FrontierFinance benchmark at an 11 times lower cost. Built on NVIDIA's Nemotron 3.5 Lightning mixture-of-experts architecture, Sterling achieved a score of 58.9% on the benchmark compared to 55.9% for Fable 5.1, while reducing the cost per question from $6.89 down to $0.61.
The model was developed by a small team using a specialized harness and a four-stage training pipeline consisting of supervised fine-tuning on synthetic and expert traces, followed by offline and online group relative policy optimization. To support the small model size, Samaya streamlined system prompts, merged tools, and introduced helper models like a condenser and a compactor to manage context and summarize traces. The development demonstrates that co-training a small model alongside its execution harness can successfully deliver domain-specific leadership in financial research tasks.
Why it matters: This release proves that smaller, highly optimized models paired with efficient harnesses can outperform larger frontier systems in specialized domains at a fraction of the cost.