Overview
Ling-3.0-flash-Fin is the first finance-enhanced model in the Ant Ling family. inclusionAI published it to Hugging Face on 3 September 2026 under the MIT license, describing it as developed by Ant Group with leading financial institutions and domain experts. It is not a new architecture: it extends Ling-3.0-flash through continued training on high-quality financial data, so it keeps the base model's 124 billion total parameters, 5.1 billion activated per token, and 256K context window.
The capabilities the model card foregrounds are end-to-end research rather than single-turn question answering. inclusionAI describes it connecting information retrieval, evidence review, calculation, modelling and report preparation instead of treating those as isolated tasks; prioritising authoritative sources so answers stay traceable; reconciling reporting periods, definitions, assumptions and conflicting figures across annual reports, earnings releases, regulatory filings and research materials; and working in valuation and spreadsheet workflows — formulas, actual-versus-estimate updates, cross-sheet dependencies, balance checks, scenario analysis and editable model delivery.
It was evaluated on FinFIRST, FinSearchComp Verified, FinCRAFT, Finance Agent, APEX-Agents, SpreadsheetBench and τ³-Banking, covering source-grounded retrieval, investment research, long-horizon execution, valuation modelling, spreadsheet operations and banking workflows. inclusionAI reports it as competitive with both similarly sized models and substantially larger general-purpose ones, with particular strength in source selection and tool-intensive financial tasks; the results are published as a chart image rather than a numeric table, so individual scores are not transcribed here. FinFIRST, the source-grounded financial search benchmark, is open-sourced alongside the model so those claims can be checked independently.
Deployment is unchanged from the base model. The checkpoint is released in BF16 and, because it shares Ling-3.0-flash's architecture, it runs on the same SGLang and vLLM runtimes with the base model's deployment guide. Thinking mode is on by default, and inclusionAI recommends temperature 1.0, top-p 0.95 and top-k 20 for general inference. The model card is explicit about the limits: as a first finance-enhanced release it still needs validation in complex long-horizon workflows, key assumptions and valuation results require professional review, and its output does not constitute investment advice. inclusionAI says future releases will explore finance-enhanced models at larger scales.
| Released | 2026-09-03 |
|---|---|
| License | MIT |
| Weights | Open weights |
| Parameters | 124B total · 5.1B activated per token (Mixture-of-Experts) |
| Context | 256K |
| Architecture | Hybrid-linear Mixture-of-Experts, inherited unchanged from Ling-3.0-flash |
| Modalities | Text |
| Status | Generally available — open weights on Hugging Face |
Strengths
- MIT-licensed open weights — downloadable, self-hostable, fine-tunable and usable commercially
- Financial-domain continued training on top of Ling-3.0-flash, with no change to the architecture or the cost profile
- Activates only 5.1B of 124B parameters per token, so domain capability does not come with dense-model serving cost
- Source-grounded search: the model card emphasises prioritising authoritative sources for traceable answers, and open-sources the FinFIRST benchmark to evaluate it
- Multi-document reasoning across annual reports, earnings releases, regulatory filings and research materials, reconciling periods, definitions and conflicting figures
- Drops into existing Ling-3.0-flash deployments — same SGLang and vLLM runtimes, same launch recipe
Best for
- Long-horizon financial research agents that retrieve, review evidence, calculate and draft a reviewable report
- Valuation and spreadsheet workflows involving formulas, cross-sheet dependencies, balance checks and scenario analysis
- Reconciling figures across filings and research materials where reporting periods and definitions disagree
- Self-hosted finance assistants at institutions that cannot send documents to a hosted API
How to access
| Provider | Model ID |
|---|---|
| OpenRouter ↗ | inclusionai/ling-3.0-flash-fin |
FAQ
How does Ling-3.0-flash-Fin differ from Ling-3.0-flash?
It is the same model architecture continued-trained on high-quality financial data. Total and activated parameter counts (124B / 5.1B), the 256K context window and the MIT license are unchanged; the difference is domain capability in financial research, valuation and banking workflows.
Is it open source?
The weights are published on Hugging Face under the MIT License, so you can download, self-host, fine-tune and use the model commercially.
How do I serve it?
The checkpoint is released in BF16 and shares Ling-3.0-flash's architecture, so the same SGLang and vLLM runtimes work — inclusionAI points to the Ling-3.0-flash deployment guide. Thinking mode is enabled by default, with temperature 1.0, top-p 0.95 and top-k 20 recommended for general inference.
What was it evaluated on?
FinFIRST, FinSearchComp Verified, FinCRAFT, Finance Agent, APEX-Agents, SpreadsheetBench and τ³-Banking — covering source-grounded retrieval, investment research, long-horizon execution, valuation modelling, spreadsheet operations and banking workflows. inclusionAI open-sourced the FinFIRST dataset alongside the model.
Can its output be used as investment advice?
No. inclusionAI states that key assumptions, valuation results and investment conclusions require professional review and do not constitute investment advice, and that this first finance-enhanced release still needs validation in complex, long-horizon workflows.