Samaya AI released Sterling, a 30B-parameter mixture-of-experts model with 3B active parameters post-trained via LoRA on NVIDIA's Nemotron 3.5 Lightning for agentic financial research. On the 220-query FrontierFinance benchmark scored by a three-model judge panel, Sterling achieves 58.9% at $0.61 and 282 seconds per question, outperforming Claude Fable 5.1's 55.9% at $6.89 and 268 seconds—an 11x cost reduction. High-effort mode raises accuracy to 60.6% at $1.83. The base Nemotron model scores 42.5%. Training proceeded in four stages on Thinking Machines Lab's Tinker API: SFT on 968 synthetic traces generated by Kimi K3 with LLM-judge selection; SFT on best-of-8 strategies across 1,819 expert questions (filtered to 1,157), where best-of-8 met 80% of rubric items versus 66% for single best; offline GRPO scoring all eight teacher answers; and online GRPO with eight model generations per question using truncated importance sampling to correct MoE sampler mismatch. The production harness was adapted for the small model by stripping detailed guidance, merging 37 tools to shorten prompts, and adding a condenser for tool results, a trace compactor, and a research budget cap before handoff to a separate report writer. Median tool calls rose from 3 to 17. On the hardest third of questions, accuracy improved from 31.8% to 45.2%. All results are self-reported; no public weights or API are announced.