Samaya Sterling a tiny AI finance agent beats a bigger one at a fraction of the cost

How it works

A program that plans financial research, calls tools to gather data, and writes reports automatically.

1. 🧠Plan the research and pick which tools to call
2. 🔧Call up to 37 finance tools to gather evidence
3. ✍️Summarize findings and write a full research report

What you could do with it

• A finance researcher or investment analyst could conduct end-to-end financial research—planning studies, calling 37 finance tools, gathering evidence, and writing reports—using a small AI agent built by Samaya AI ².
• A developer or company could deploy an affordable financial research agent for about 11 times less money than using Claude Fable 5.1, making research tools cheaper to run at scale ¹.
• An investment team could improve screening and macro research performance, moving from 31.8 percent on the hardest questions to 45.2 percent after post-training ¹.

An exampleInvented to explain it

Maya reviews stocks from her small apartment desk

MayaFind everything about NovaTech's latest revenue report

SterlingSearching 37 finance tools for evidence. I will call them now.

MayaHow long will this take and what will it cost?

SterlingPlanning the study, calling tools, gathering evidence, then writing a report. Much less than big competitors.

Sterling completed the full research workflow and delivered a report at a fraction of the usual cost.

A smaller model doing the job of a larger one

Samaya AI has released Sterling, a model designed to act as a financial research agent. It is based on Nemotron 3.5 Lightning, a model from NVIDIA that has 30 billion total parameters but only uses 3 billion at any one time — a design called a mixture of experts, where different parts of the model handle different tasks ¹².

On Samaya's own FrontierFinance benchmark, which tests AI ability in real investment workflows, Sterling scores 58.9 percent accuracy ¹. That beats Claude Fable 5.1's score of 55.9 percent . The main difference, though, is cost. Each research question costs 61 cents for Sterling, compared with $6.89 for Fable 5.1 — roughly 11 times cheaper ².

The model itself only costs about 4 cents per question; the rest covers the tools and helper models it uses ¹. Samaya says every training stage added quality, moving from 42.5 percent accuracy on the base model to 58.9 percent after four stages of training . On the hardest third of questions, performance rose from 31.8 percent to 45.2 percent .

How Sterling works

A financial research agent is a program that plans research, calls tools to get data, gathers evidence, and then writes a report ². Sterling does this by first being trained with supervised learning to call 37 different finance tools with correct inputs, then learning from improved strategies .

Samaya adapted its production finance tool system for a small model. They simplified the instructions, merged some tools to shorten prompts, and added two helper models to manage context — a condenser that shortens long tool results, and a compactor that summarizes the research trace as it grows ². A research budget limits how long the agent works before handing findings to a separate report writer .

Training used LoRA, a method that adjusts only part of a model's parameters to save time and money, across four stages on Kimi K3 as the teacher model ¹². The first stage used 968 synthetic questions. The second used 1,819 expert-written questions with eight different research strategies, keeping only the best answer per question . Two further stages used reinforcement learning, first offline scoring existing answers, then online where the model answered each question eight times and learned from graded attempts .

What is still unknown

Samaya says no benchmark question was used in training, and checkpoints were chosen on 100 internal validation questions ¹. However, all results are self-reported by Samaya on their own benchmark . The model does not have public weights or an API announced yet; the base model runs on a single H100 GPU with 80 GB memory at 256K context, but Sterling's serving hardware and context are not disclosed . No figures are reported for other benchmarks beyond FrontierFinance.

The broader takeaway is that co-training a small model with its tool system can narrow the gap with larger models in specialized domains ².

How good is it

On Samaya's own FrontierFinance benchmark, which tests AI ability in real investment workflows using 220 public queries scored against expert-written rubrics, Sterling scores 58.9 percent accuracy ¹. That beats Claude Fable 5.1's score of 55.9 percent on the same test . Every training stage added quality, moving from 42.5 percent on the base model up to 58.9 percent after four stages . All results are self-reported by Samaya AI .

accuracy on real investment research workflows. Higher is better. Numbers: Samaya AI (self-reported). Chart drawn by Herald of AI.

Chart from the announcement: The chart compares Samaya Sterling against other AI systems on financial research performance, where a taller curve means a higher score and better results

Picture: samaya.ai

Chart from the announcement: The chart compares model scores on the FrontierFinance benchmark, where a taller bar means better performance

Picture: samaya.ai

Chart from the announcement: The chart compares cost-per-question for financial research agents, where a shorter bar means lower cost and better value

Picture: samaya.ai

Key facts

• Sterling scores 58.9% on FrontierFinance, beating Claude Fable 5.1's 55.9%
• Cost per question drops 11x: $0.61 vs $6.89 for Fable 5.1
• Built on NVIDIA's Nemotron 3.5 Lightning with 30B total / 3B active parameters
• Trained in 4 stages using LoRA with Kimi K3 as teacher model
• No public weights or API announced; base model fits on single H100 80GB

Sources

1. Samaya AI Releases Sterling: A 3B-Active Finance Research Agent Model that Beats Claude Fable 5.1 on FrontierFinance at — marktechpost.com
2. Samaya Sterling: A post-trained model for state-of-the-art financial research | Samaya