Evidence at a glance
The Research Workflow Matters as Much as Model Size
Samaya AI has released Sterling, a post-trained agent model for financial research. It is built on NVIDIA’s Nemotron 3.5 Lightning, a mixture-of-experts model with 30 billion total parameters and 3 billion active parameters. Sterling is designed to plan research, call financial tools, gather evidence, and pass its findings to a separate report writer.
The release is worth a technical lead’s attention not because it simply proves that a small model beats a large one, but because the comparison is about more than the model. Samaya tuned the model alongside its research workflow and tool-calling harness, then reported a higher score at lower cost on its own FrontierFinance benchmark. The more useful question is whether, for bounded tasks, some capabilities can move from a general-purpose model into a deliberately designed execution system.
Moving Complexity Out of the Prompt
Sterling’s harness does not work by stuffing more instructions into a small model’s prompt. Samaya says it adapted its production finance research framework by removing detailed guidance from the system prompt and tool descriptions, and merging tools to shorten the prompt. Two helper models manage context: one condenses long tool results, while another summarizes the trace as it grows. A research budget limits how long the agent can work, after which a report writer answers from the material gathered.
This design shifts some of the burden of remembering rules and organizing history from the core model to the surrounding system. It also means that model inference alone is not the right cost measure. Sterling’s model costs about $0.04 per question, while the reported total of $0.61 includes tools and helper models. The lower price is therefore a property of the configured workflow, not simply a consequence of having fewer active parameters.
Training Breaks “Doing Research” into Stages
Samaya reports a four-stage training process using LoRA through the Tinker training API. The first stage used supervised fine-tuning on synthetic traces: the team generated 968 analyst questions, had Kimi K3 answer each twice, and used a language-model judge to select the better response. The model also learned to call 37 tools with valid arguments. In the second stage, the teacher tried eight research strategies on 1,819 expert-written questions, and the highest-graded answer was retained. Filtering left 1,157 questions. Samaya says the best of eight strategies covered 80% of essential rubric items, compared with about 66% for a single best strategy.
The final stages were offline and online GRPO. The offline stage compared existing sets of eight teacher answers without generating new content at each step. In the online stage, Sterling produced eight attempts per question and learned from their grades. The team also used truncated importance sampling to address a mismatch involving the MoE sampler. The point is not any one training trick, but the sequence: teach tool use and research strategies, then improve the model through its own attempts, making the workflow itself part of what is trained.
A Lead on the Scoreboard, with Important Caveats
On FrontierFinance’s 220 public questions, Samaya reports a score of 58.9% for Sterling and 55.9% for Claude Fable 5.1. The reported cost per question is $0.61 for Sterling versus $6.89 for Fable, while their average runtimes are 282 and 268 seconds respectively. Sterling’s high-effort mode reaches 60.6% at $1.83 per question and 372 seconds. On the hardest third of the questions, Sterling rises from the base model’s 31.8% to 45.2%.
These figures show trade-offs among quality, cost, and speed rather than an across-the-board win: high-effort mode scores better but costs more and takes longer, while Sterling’s standard mode is slightly slower than Fable. A panel of three models grades answers against expert-written rubrics. Samaya says no benchmark questions were used in training and that checkpoints were selected using 100 internal validation questions. But Samaya built FrontierFinance and the harness, and ran the Sterling evaluation, so the results need to be read with benchmark ownership in mind. The reported 45.2% on the hardest third also leaves substantial room for improvement on difficult questions.
A Case for Testing Specialist Workflows, Not Immediate Deployment
For technical teams, Sterling’s most useful lesson is not to copy a particular model or training recipe, but to account for the parts of an agent’s performance: research planning, tool selection, context compression, execution budgets, and final writing. Samaya reports stage-by-stage scores rising from 42.5% for the base model to 51.4%, 54.8%, 57.5%, and finally 58.9%. Median tool calls also rose from three to 17. That pattern suggests that better scores came with more intensive tool use, so cost analysis needs to cover the full execution trace.
The base Nemotron model’s specifications should not be mistaken for Sterling’s deployment requirements. The material says the base model fits on one H100 80GB in BF16 at a 256K context, but Sterling’s serving context and hardware have not been disclosed. The base model’s context limit of up to one million tokens does not establish that Sterling is served at that length. Samaya has not announced public Sterling weights or an API. For now, the sound judgment is to treat Sterling as a useful specialist-agent case study: validate quality and full-workflow cost on your own tasks before considering adoption, rather than extrapolating production value from one benchmark’s cost-performance result.