Another post-trained small model in a vertical field, Samaya Sterling, has achieved a 3.0 percentage point improvement on the FrontierFinance dataset, outperforming the top-ranked Frontier model, Claude Fable 5.1, while reducing the cost per question by 11 times ($0.61 vs. $6.89). They trained the model on Nemotron 3.5 Lightning using Thinking Machines Lab's training API Tinker via LoRA, with four phases: two phases of supervised fine-tuning (SFT) and two phases of reinforcement learning (RL). After training, the model's accuracy on the FrontierFinance dataset improved from 42.5% to 58.9% (see figure). Samaya AI (http://samaya.ai) is an AI company focused on the financial services sector, providing expert-level AI intelligence platforms/AI agents for investment professionals (buy-side/sell-side). Founders/Team: The founder and CEO is Maithra Raghu (PhD), who previously conducted research at Google Brain. Business Model: Providing AI agents for asset management, equity research, investment banking, and wealth management for tasks such as research, meeting preparation, and argument review. Clients include large institutions such as Morgan Stanley (across research, sales and trading, and investment banking divisions); claims thousands of users in production environments with rapidly growing adoption. Funding: $43.5 million, with participants including former Google CEO Eric Schmidt, Meta Chief AI Scientist Yann LeCun, Two Sigma co-founder David Siegel, and former Goldman Sachs executive/current Sixth Street Vice Chairman Marty Chavez. Technical Report: https://samaya.ai/blog/rl-post-trained-agentic-model