Samaya AI has presented Sterling, a post-trained model for agentic financial research that, on the FrontierFinance benchmark, outperforms Claude Fable 5.1 while spending about eleven times less per question. The result is self-reported by the company on its own benchmark: Sterling scores 58.9% against 55.9% for Fable 5.1, at $0.61 per question versus $6.89.
Sterling is built on Nemotron 3.5 Lightning, an Nvidia model with 30 billion total parameters in a mixture-of-experts architecture, of which 3 billion are active per token. The base model can run on a single 80 GB H100 in BF16 with up to 256K tokens of context. Samaya has not disclosed Sterling's serving context or hardware.
Image: samaya.ai — https://samaya.ai
An architecture built for a small model
The core of the work is adapting the software harness that Samaya uses in production for finance. Detailed instructions were removed from the system prompt and tool descriptions, and some tools were merged to shorten the prompt. Two helper models manage context: a condenser shortens long tool results, and a compactor summarizes the trace as it grows. A research budget caps how long the agent works, and a separate model writes the final report from what was found.
Four training stages
Training used LoRA on Thinking Machines Lab's training API, in four steps. First, Samaya generated 968 synthetic analyst questions, each answered twice by the teacher model Kimi K3, and an LLM judge picked the better answer, so the model learned to call 37 tools with valid arguments. Second, Kimi K3 answered 1,819 expert-written questions using eight research strategies; the best-graded answer per question became training data, leaving 1,157 questions after filtering. Two GRPO reinforcement-learning stages followed: an offline one that compares the eight teacher answers without generating new ones, and an online one in which the model answers each question eight times and learns from the graded attempts.
Samaya states that no benchmark question was used in training and that checkpoints were chosen on 100 internal validation questions. Claude Fable 5.1, launched in September, has slashed cache costs by 75%, the figure Sterling uses to compare its own cost per prompt.
The numbers
FrontierFinance has 220 public questions scored against expert-written rubrics, with grading by a panel of three models. Sterling takes on average 282 seconds per question, against 268 for Fable 5.1. Each training stage added quality: from 42.5% for the base model to 51.4%, then 54.8%, 57.5% and finally 58.9%. The median number of tool calls rose from 3 to 17. In high-effort mode Sterling reaches 60.6%, at $1.83 and 372 seconds per question. On the hardest third of questions it rose from 31.8% to 45.2%.
The model alone costs $0.04 per question; the rest of the $0.61 goes to tools and helper models.
What is not known
Sterling has no public weights and no announced API, unlike the Nvidia base model, which is open-weight under the OpenMDW 1.1 license. Sterling's token prices have not been disclosed. The comparison with Fable 5.1 rests on Samaya's own benchmark, published in July to measure finance agents across the whole investor workflow, from screening to modelling.
Samaya's previous result is recent: in July the company said its agent system scored 50.8% on FrontierFinance, beating the best frontier model at four times lower cost. Sterling lifts that score to 58.9% with a far smaller model than the original system.
The lesson Samaya draws for developers is practical: co-training a small model with the harness that uses it can make it win in a narrow domain. The result remains self-reported and covers a single benchmark.
No comments yet. Be the first!