Samaya AI has released Samaya Sterling, a post-trained model for agentic financial research. Sterling is built on NVIDIA’s Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters. On Samaya’s FrontierFinance benchmark, it scores 58.9%. Claude Fable 5.1 scores 55.9%. Each Sterling question costs $0.61, against $6.89 for Fable 5.1. For developers, the takeaway is a concrete recipe: co-train a small model with its harness, and it can win in a narrow domain.
TL;DR
- Size: 30B total, 3B active (MoE). Base model context: up to 1M tokens. Sterling’s serving context: not disclosed.
- Runs on: the base model fits a single H100 80GB in BF16 (256K context). Sterling’s serving hardware: not disclosed.
- Performance: beats every high-quality system on the FrontierFinance leaderboard on quality per dollar (self-reported).
- Best: 58.9% on FrontierFinance (60.6% in high-effort mode), vs 55.9% for Claude Fable 5.1.
- Worst: 45.2% on the hardest third of FrontierFinance questions.
What is Samaya Sterling?
Samaya Sterling is a LoRA post-trained version of Nemotron 3.5 Lightning, tuned to work as a financial research agent. It plans research, calls finance tools, gathers evidence and hands findings to a separate report writer. According to Maithra Raghu’s X announcement, it started as a 2-person effort on agentic orchestration efficiency.
How does the harness support a 3B-active model?
Samaya team adapted its production finance harness for a small model. Detailed guidance was stripped from the system prompt and tool descriptions. Tools were merged to shorten the prompt. Two helper models manage context. A condenser shortens long tool results. A compactor summarizes the trace as it grows. A research budget caps how long the agent works. After that, the report writer answers from what was found.
How was Sterling trained?
Training used LoRA on Tinker, Thinking Machines Lab’s training API. It ran in 4 stages:
- SFT on synthetic traces: Samaya generated 968 synthetic analyst questions. The teacher, Kimi K3, answered each twice, and an LLM judge picked the better answer. The model learned to call 37 tools with valid arguments.
- SFT on the best of 8 strategies: Kimi K3 answered 1,819 expert-written questions using 8 research strategies. The best-graded answer per question became training data. Best-of-8 met 80% of essential rubric items, against about 66% for the single best strategy. After filtering, 1,157 questions remained.
- Offline GRPO: GRPO scored all 8 teacher answers against each other. No new generation was needed, so each step was cheap.
- Online GRPO: The model answered each question 8 times and learned from graded attempts. Truncated importance sampling corrected a known MoE sampler mismatch.
How does Sterling score on FrontierFinance?
FrontierFinance has 220 public queries scored against expert-written rubrics. The grader uses a 3-model judge panel. Samaya team states that no benchmark question was used in training. Checkpoints were chosen on 100 internal validation questions.
Sterling scores 58.9% at $0.61 and 282 s per question. Fable 5.1 scores 55.9% at $6.89 and 268 s. Samaya’s own high-effort system scores 2.9 points lower at $1.81.
Every stage added quality: 42.5% (base), 51.4%, 54.8%, 57.5%, then 58.9%. Median tool calls rose from 3 to 17. In high-effort mode, Sterling reaches 60.6% at $1.83 and 372 s. On the hardest third, it rose from 31.8% to 45.2%.
The model itself costs $0.04 per question. Tools and helper models make up the rest of the $0.61.
‘;stg.appendChild(c)});
var cols=stg.querySelectorAll(‘.col’);
function stgNext(){if(stgI>=5)return;var c=cols[stgI];c.classList.add(‘shown’,’lit’);c.querySelector(‘.bar’).style.height=((S[stgI][0]-30)/32*150+20)+’px’;cols.forEach(function(x,k){if(k
document.getElementById(‘mtpStgNext’).onclick=stgNext;
document.getElementById(‘mtpStgReset’).onclick=function(){stgI=0;cols.forEach(function(c){c.classList.remove(‘shown’,’lit’);c.querySelector(‘.bar’).style.height=”0″});document.getElementById(‘mtpStgInfo’).textContent=””;stgNext();};
var D=[[“Sterling, high effort”,60.6,1.83,372,1],[“Samaya Sterling”,58.9,0.61,282,1],[“Samaya System, high effort”,56.0,1.81,278,0],[“Claude Fable 5.1”,55.9,6.89,268,0],[“Nemotron base”,42.5,null,null,0]];
var notes={q:”Share of rubric items satisfied on 220 public FrontierFinance queries. Samaya System figure derived from the blog (2.9 points below Sterling).”,c:”Cost per question includes the model, tools and helper models. Lower is better. Base model cost not disclosed.”,l:”Seconds per question. Lower is better. Base model latency not disclosed.”};
var mode=”q”;
function drawBars(m){mode=m;var w=document.getElementById(‘mtpBars’);w.innerHTML=”;var idx=m===’q’?1:m===’c’?2:3,max=m===’q’?70:m===’c’?7.5:400;
D.forEach(function(d){var v=d[idx];var r=document.createElement(‘div’);r.className=”row”;var lab=v===null?’n/a’:(m===’q’?v.toFixed(1)+’%’:m===’c’?’$’+v.toFixed(2):v+’ s’);r.innerHTML=’
‘+d[0]+’
‘+lab+’
‘;w.appendChild(r);var f=r.querySelector(‘.fl’);setTimeout(function(){f.style.width=v===null?’0′:(v/max*100)+’%’},40)});
document.getElementById(‘mtpBarNote’).textContent=notes[m];post();}
R.querySelectorAll(‘#mtpSeg button’).forEach(function(b){b.addEventListener(‘click’,function(){R.querySelectorAll(‘#mtpSeg button’).forEach(function(x){x.classList.toggle(‘on’,x===b)});drawBars(b.getAttribute(‘data-m’))})});
R.querySelectorAll(‘.cd’).forEach(function(c){c.addEventListener(‘click’,function(){c.classList.toggle(‘open’);setTimeout(post,450)})});
window.addEventListener(‘load’,post);window.addEventListener(‘resize’,post);setTimeout(post,300);
})();
