Samaya's 30B post-trained model beats frontier on financial research at 11x lower cost
maithra_raghu · x · 2026-10-29
Samaya introduced Sterling, a financial research model post-trained from Nemotron 3.5 Lightning (30B MoE, 3B active params) that scores 58.9% on FrontierFinance vs 55.9% for the best frontier model, at $0.61 vs $6.89 per task — an 11x cost reduction.
The recipe has four stages: two SFT phases (learning to use financial tools, then to build agentic research strategies), offline GRPO distilling from eight diverse teachers, and online GRPO for end-to-end trial-and-error. Co-adaptation of model and harness — better tool use and compaction — let the small model handle longer, deeper research. Follow-ups include multi-agent research swarms, firm-specific models, and a larger model showing even greater gains.
More from coding & agent
- Jason Kneen shares slop_eisley 0.6.22 roadmap for his AI-built FPS game — jasonkneen · 2026-10-29
- Free Oct 29 Every livestream: Context Engineering With DSPy, open to all — danshipper · 2026-10-29
- Open-source lead gen agent Jev at $0.0024/lead; blogger ran a job hunt for $0.13 — JaynitMakwana · 2026-10-29
- Hands-on exercises and resources for GitHub's new Copilot features — jennifermarsman · 2026-10-29
- YC open-sources QM agent harness; E2B adds one-command sandbox deploy — badphilosopher · 2026-10-29
- Ploid launches a living index of 2 billion people built for AI agents — ycombinator · 2026-10-29