Samaya introduced Sterling, a post-trained Nemotron 3.5 Lightning model with 30 billion total and 3 billion active parameters, using two supervised and two reinforcement-learning stages. On 220 public FrontierFinance questions, it reports satisfying 58.9% of rubric items at $0.61 total cost per question, compared with Claude Fable 5.1’s 55.9% and $6.89. The research model accounts for $0.04 of Sterling’s cost; tools and helper models account for the rest. These are developer-reported results from jointly adapting financial tools and context compression, not evidence of investment returns or superiority on other tasks.