Samaya AI has released Samaya Sterling, a post-trained model built for agentic financial research — and it beats Claude Fable 5.1 on Samaya’s FrontierFinance benchmark while costing roughly 11x less per question. The kicker: Sterling runs on just 3B active parameters.
Samaya Sterling (Credits: Samaya AI)
Samaya Sterling vs Claude Fable 5.1 on FrontierFinance
Sterling is a LoRA post-trained version of NVIDIA’s Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters. On the FrontierFinance benchmark — 220 public queries scored against expert-written rubrics by a 3-model judge panel — Sterling scores 58.9% (60.6% in high-effort mode), versus 55.9% for Claude Fable 5.1. Each Sterling question costs $0.61, against $6.89 for Fable 5.1. The model itself accounts for just $0.04 per question; the rest goes to tools and helper models.
Latencies are comparable: 282 seconds per question for Sterling against 268 seconds for Fable 5.1. Every training stage added quality, climbing from 42.5% at the base model through 51.4%, 54.8% and 57.5% to the final 58.9%, while median tool calls per question rose from 3 to 17.
How Samaya Sterling was trained
According to Samaya’s breakdown, Sterling started as a 2-person effort on agentic orchestration efficiency, with training across four stages on Tinker, Thinking Machines Lab’s training API. First came SFT on synthetic traces: Samaya generated 968 synthetic analyst questions, teacher model Kimi K3 answered each twice, and an LLM judge picked the better answer. Then SFT on best-of-8 strategies: Kimi K3 answered 1,819 expert-written questions using eight research strategies, keeping the best-graded answer per question. Offline GRPO then scored all eight teacher answers against each other without new generation, followed by online GRPO where the model answered each question eight times and learned from graded attempts.
Engineering mattered as much as training. Samaya stripped detailed guidance from the system prompt and tool descriptions, merged tools to shorten prompts, and used two helper models — a condenser that shortens long tool results and a compactor that summarizes the trace as it grows — plus a research budget that caps how long the agent works before a separate report writer answers from what was found. The base model context reaches up to 1M tokens, and the model fits on a single H100 80GB in BF16.
The takeaway for developers, per the release: co-train a small model with its harness, and it can win in a narrow domain — at a fraction of the price of frontier systems.
Also Read: Hermit: Infini-AI-Lab’s Open-Source Framework Trains AI Agents Without Touching the Model
Source: MarkTechPost — read the full story there.
