Overview
Samaya is a research platform for investment teams on both the buy and sell side, with Morgan Stanley's research division among its partners. Sterling is the finance research model post-trained for that platform: an NVIDIA Nemotron 3.5 Lightning base with 30B total and 3B active parameters, further trained with LoRA on financial material, built to connect a firm's own knowledge with filings, earnings transcripts and licensed data and answer with sources attached.
The platform side is organised around the research workflow. An assignment can run now, run on a schedule, or fire when a connected source publishes something, and each run returns a report with citations. Every finding leads back to the passage behind it, labelled as first-party, third-party or web material. During earnings season it can follow the call transcript live, compare the quarter against the last one and hold a watchlist of your own follow-up questions.
There are two ways in. MCP brings earnings recap, model build and fact checker skills into Claude and ChatGPT so they can be called where you already write, while the API exposes cited answers and financial retrieval to internal tools. Access follows the source permissions a firm already has in place.
Key Features
- Finance post-training: LoRA training on top of Nemotron 3.5 Lightning concentrates reasoning budget on earnings interpretation, like-for-like comparison and thesis testing.
- Traceable citations: Follow a conclusion down to the source passage, with first-party, third-party and web references labelled separately for review.
- Scheduled research agents: The same assignment can run on demand, on a schedule or when a source publishes an event, with a view of what is active, paused or needs attention.
- Live earnings coverage: Follow the call transcript in real time, track the quarter's themes, compare management commentary with the prior quarter and add your own questions to watch.
- MCP skills: Run earnings recap, model build and fact checker from Claude or ChatGPT without switching back to the platform.
- API and governed access: Cited answers and financial retrieval plug into enterprise tools, with source permissions and entitlements carried along with the research.
Use Cases
- Covering a company, put the quarter's filings, the earnings call and sell-side notes side by side to find the number that actually moved.
- Run a thesis check before an investment committee and pull the source passage behind each assumption.
- Compare internal memos against external disclosures to see which calls new data has overtaken.
- Draft an earnings recap inside Claude, then return to the platform to verify each citation.
- Prepare talking points for advisors that trace company developments through to model portfolios.
- Turn peer comparison into a scheduled job that refreshes daily through earnings season.
Pros
- Post-training on financial material brings model cost for the same task to roughly a tenth of comparable frontier models.
- Scores 58.9 on the in-house FrontierFinance benchmark, with the answers and citations supplied together.
- First-party, third-party and web sources are labelled separately, making the research auditable.
- Research agents support schedules and event triggers, covering ongoing tracking rather than one-off prompts.
- MCP and API give two integration paths into existing writing and research tools.
- Access follows source permissions, with SOC 2 and deployment options open to negotiation.
Pricing
Priced per institutional agreement. No subscription price is published, so conversations start with sales. Cost structure depends on usage and access route, with seats, API calls and MCP skills priced separately in the contract. Deployment can be arranged to run research inside existing permission structures, and SOC 2 reporting plus data-handling terms come with the agreement.
Summary
The hard part of finance research is rarely whether the model is clever enough. It is whether the answer is right, whether you can trace it, and whether the spend is defensible. Sterling takes the third path: a 30B MoE base with finance post-training brings model cost for a single research task to about a tenth of comparable frontier models, and every conclusion stays tied to a source passage. For teams that ship reports daily, a citation you can check beats a few more points of accuracy.
Version History
- Samaya Sterling listed: finance research agent model (2026-10-31): Form at listing: a finance research model LoRA post-trained on NVIDIA Nemotron 3.5 Lightning (30B total, 3B active parameters), scoring 58.9 on the in-house FrontierFinance benchmark at roughly a tenth of the model cost of comparable frontier models. The platform provides cited research agents that run on demand, on schedule or on source events, follow earnings calls live, and trace each conclusion to a labelled source passage. Access via MCP skills or API, permissions carried from the source system, SOC 2 with negotiable deployment.