Samaya AI has introduced Samaya Sterling, a post-trained model for agentic financial research, meaning it can autonomously chain together research steps and tool calls. According to the company, it scores 58.9% on the FrontierFinance benchmark, compared to 55.9% for Claude Fable 5.1, at a cost per question of $0.61 versus $6.89.
Sterling is based on NVIDIA Nemotron 3.5 Lightning, a 30B parameter Mixture of Experts (MoE) model (an architecture that only activates a portion of its parameters per request). Only 3B parameters are active. According to Maithra Raghu’s announcement on X, the project started as a two-person effort on the efficiency of agentic orchestration.
A harness adapted for a small model
The harness refers to the set of instructions, tools, and mechanisms that frame the model. Samaya adapted its finance harness, used in production, for this smaller model. Detailed instructions were removed from the system prompt and tool descriptions, and some tools were merged to shorten the prompt. Two auxiliary models manage context: a condenser shortens long tool results, while a compactor summarizes the work trace as it lengthens.
A research budget finally limits the agent’s working duration. The report writer then responds based on the elements found.
Training in four steps
The training used the LoRA technique, which slightly adapts an existing model, on Tinker, the training API from Thinking Machines Lab. The first step consists of supervised fine-tuning (SFT) on 968 synthetic analyst questions. The teacher, Kimi K3, answered each question twice, and an LLM-based judge selected the best answer. The model learned to call 37 tools with valid arguments.
The second step uses 1,819 questions written by experts, processed with eight research strategies. The best answer per question served as training data, and 1,157 questions were kept after filtering. According to Samaya, this selection covers 80% of the essential criteria on the grid, compared to about 66% for the best single strategy. The third and fourth steps are two phases of reinforcement learning using the GRPO method: the first, offline, compares the eight teacher answers against each other; the second, online, has the model answer each question eight times.
Self-reported results on FrontierFinance
The FrontierFinance benchmark includes 220 public queries, scored according to grids written by experts, using a panel of three judge models. Samaya states that no benchmark questions were used in training, and that checkpoints were selected based on 100 internal validation questions.
Sterling achieves 58.9% in 282 seconds per question, compared to 268 seconds for Claude Fable 5.1. The Samaya system in ‘high effort’ mode scores 56.0%, at $1.81 per question. The score improves at each step, from 42.5% for the base model to 51.4%, 54.8%, 57.5%, then 58.9%. The median number of tool calls increases from 3 to 17. On the most difficult third of questions, the score rises from 31.8% to 45.2%. In ‘high effort’ mode, Sterling reaches 60.6% for $1.83 and 372 seconds. The model itself costs $0.04 per question, with the rest of the cost coming from tools and auxiliary models. The base model, according to the source, fits on a single 80GB H100 GPU in BF16, with a 256K token context. Its context window can reach 1M tokens. Samaya does not disclose the context window, the serving hardware for Sterling, and no public weights or API have been announced for it. All results are self-reported and published on Samaya’s benchmark.
