Skip to content
Neural Daily

ModelsNews

Samaya Sterling: 3B active params, beats Claude Fable 5.1

🇭🇺 Olvasd el magyarul →

Samaya post-trained a 30B Nemotron model with 3B active parameters for financial research. It scores 58.9% on Samaya's FrontierFinance benchmark, above Claude Fable 5.1's 55.9%, at roughly an eleventh of the cost.

By: Mikola Máté Patrik · · 2 min read

Line chart showing Samaya Sterling rising from 42.5% to 58.9% on FrontierFinance across training stages, above Claude Fable 5.1
Image: Samaya · Press kit · Official social image from the Samaya research post · Source

What's new

In an October 28 research post, Samaya introduced Sterling, a model post-trained for financial research (Samaya). It starts from NVIDIA's open Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B parameters active at a time. On Samaya's FrontierFinance benchmark, which covers real analyst workflows across the investment process, the base model improved from 42.5% to 58.9% (Samaya).

For comparison, the best frontier model, Claude Fable 5.1, scores 55.9% at $6.89 per task. Sterling beats it at $0.61, about an 11x cost reduction. FrontierFinance questions were strictly held out from training, and checkpoints were selected on a separate pool of 100 internal validation questions.

How it was trained

  • Two SFT stages: first on 968 synthetic analyst questions, then on expert-written and expert-graded tasks, using traces from eight different teacher models.
  • Two RL stages: offline GRPO, then online RL where the model learns from its own mistakes.
  • Training used LoRA on Thinking Machines Lab's Tinker API.
  • The harness was adapted for a small model: merged tools, a condenser for long tool results, a compactor for the growing trace and a research budget after which a separate report writer answers.

Why it matters

A key claim is that the gains come from co-adapting model and harness: the model learns to use the finance tools during training, and context compression lets a small model sustain long research runs. A notable detail: adding a context-length penalty to the reward cut cost by 30% for a 1.1-point quality loss, making the quality and cost trade-off tunable at training time. One caveat is that FrontierFinance is Samaya's own benchmark, with no independent verification yet.

Samaya says it sees even bigger gains from a larger model trained with the same recipe and plans to share those results later (Samaya). MarkTechPost also covered the announcement (MarkTechPost).

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles