Back home

Samaya Sterling

x+web 0.85
Visit source

Intro

Samaya Sterling is a post-trained model for agentic financial research released by Samaya AI on October 29, 2026. It is built on NVIDIA's open Nemotron 3.5 Lightning (a 30B mixture-of-experts model with only 3B active parameters) and was trained together with Samaya's proprietary finance-tool harness.

On Samaya's open FrontierFinance benchmark (220 expert-rubric investment questions) Sterling tops the leaderboard at 58.9%, ahead of Claude Fable 5.1 (55.9%), while costing about 11x less per question.

Blog: samaya.ai/blog/rl-post-trained-agentic-model · Benchmark: FrontierFinance

Explain

Key points

  1. Base model: NVIDIA Nemotron 3.5 Lightning, 30B MoE / 3B active, thinking enabled; trained with LoRA on Thinking Machines' Tinker API.
  2. Four-stage recipe: two supervised stages (learning from eight teacher models how to research with finance tools and aggregate citable answers) followed by offline RL and online RL.
  3. Model + harness co-adaptation: the harness exposes dozens of finance tools (37 per third-party listings) and compresses long traces so a small model can sustain long research; median tool calls rose from 3 to 17.
  4. Results: 42.5% → 58.9% on FrontierFinance; high-effort mode reaches 60.6%. Cost $0.61/question (the model itself only ~$0.04) vs $6.89 for Fable 5.1, at similar latency (~282 s).
  5. Coverage: MarkTechPost, NVIDIA developer relations posts and Chinese AI accounts on X.

Why it matters

Sterling is a concrete data point for the "small model + specialized harness beats frontier generalist" thesis in a high-value vertical, and a new searchable product name in finance AI.

Usage

How to use it

  1. Read the technical post on the Samaya blog for the training recipe and appendices.
  2. Compare systems on the public FrontierFinance leaderboard (score, latency, cost per query).
  3. Sterling runs inside the Samaya platform for investment workflows—screening, macro research, earnings and company-metric analysis; contact Samaya for access.
  4. To reproduce the idea yourself: start from an open small MoE (e.g. Nemotron 3.5 Lightning), build a domain tool harness, then SFT on teacher traces and finish with RL on rubric-graded tasks.