Finance is not only about resource allocation and industrial development; it also connects to the wealth and lives of countless households. Unlike general information processing, financial tasks involve rapidly changing information and complex conventions, with analysis spanning macroeconomics, industries, companies, and markets, and conclusions that must withstand scrutiny of data, logic, and compliance. Therefore, financial intelligence cannot stop at the Q&A level; it must enter real workflows: find trustworthy information, complete rigorous analysis and computation, and deliver research outputs that are reviewable and editable.
Building on long-term accumulation in fintech and artificial intelligence, Ant Group, together with leading domestic financial institutions and industry experts, jointly developed Ling-3.0-flash-Fin. From task definition, data systems, and training methods to evaluation standards and real-world scenario validation, all parties are deeply involved in the key stages of model development, driving the model to continuously align with the professional methods and practical needs of the financial industry.
As Ant Bailing's first finance-enhanced model, Ling-3.0-flash-Fin inherits the model architecture and long-context capabilities of Ling-3.0-flash, maintaining a configuration of 124B total parameters and 5.1B activated parameters. On this basis, through continued pre-training on financial corpora, domain post-training, and tool-use optimization, the model further strengthens its ability to process complex financial content such as annual reports, financial workbooks, and multiple research materials, while also taking deployment efficiency into account.
01 Oriented toward real workflows, honing core investment research capabilities
Ling-3.0-flash-Fin starts with the investment research scenario, focusing on honing core capabilities such as information retrieval, research reasoning, valuation modeling, and research report writing, so that information, evidence, computation, judgment, and document delivery can be completed collaboratively within the same task.
Enhanced professional capabilities, grounded in general intelligence
Ling-3.0-flash-Fin was tested on FinFIRST, jointly developed by Ant Group and CICC, as well as financial agent benchmarks including FinSearchComp Verified, Finance Agent, APEX-Agents, SpreadsheetBench, and τ³-Banking, covering scenarios such as financial information retrieval, investment research analysis, long-horizon financial task execution, valuation modeling, and banking applications. The results show that, compared with multiple general-purpose models, Ling-3.0-flash-Fin demonstrates strong competitiveness on the above tasks.
Evaluation notes
Unless otherwise specified, the sampling parameters for Ling-3.0-flash-Fin in all benchmark tests are set to temperature=1 and top_p=0.95. The reasoning effort of all models is set to the highest level.
- FinFIRST: A financial search benchmark jointly developed by Ant Group and CICC. All models are evaluated using a unified ReAct agent framework, equipped with Web Search, Visit, and Python tools. The benchmark will be open-sourced soon.
- FinSearchComp Verified: An internally expert-verified version of FinSearchComp, containing 145 T2/T3-level questions, whose reference answers have all been reviewed and revised by financial professionals. All models are evaluated using a ReAct agent framework equipped with Web Search, Visit, and Python tools. We use gpt-5-2025-08-07 as the judge model.
- FinCRAFT: An internal benchmark for real-world financial investment research scenarios, used to evaluate the ability of AI agents to complete end-to-end financial research and investment analysis tasks.
- Finance Agent v1.1 / Finance Agent v2: All models are evaluated using the official evaluation framework and public datasets in part. Results use the Strict-Pass criterion, meaning a task is judged as passed only if all of its scoring criteria are met. To reduce single-run variance, v1.1 scores are averaged over 10 runs, and v2 scores are averaged over 20 runs.
- APEX-Agents: The official evaluation framework is used. Among them, Claude-Opus-5, GPT-5.6, GLM-5.2, and Kimi-K3 use the scores published by Mercor; MiniMax-M3 and Hy3 use officially published scores; the scores of the remaining models come from internal testing.
- SpreadSheetBench V1/V2: All models use the Claude Code evaluation framework. The maximum number of interaction rounds is set to 120 for V1 and 300 for V2, with a timeout limit of 3 hours per task. The evaluation environment uses Claude Code 2.1.173 and LibreOffice 25.8.7, where Claude Code enables the default Skill tool and disables the Search function. Under this evaluation framework, the temperature for Ling-3.0-flash-Fin is set to 0.6.
- τ³-Banking: All models use the evaluation scores published by Artificial Analysis.
The improvement in professional capabilities is also inseparable from solid general capabilities. Building on Ling-3.0-flash, Ling-3.0-flash-Fin further introduces large-scale, high-quality financial domain data for continued training. While financial domain capabilities are strengthened, the model's score on AA Intelligence Index v4.1.1 rises from 38 to 41, achieving coordinated improvement in both general and professional capabilities.
02 Information retrieval: from "finding" to accuracy, completeness, and traceability
Search is the starting point of financial research. Reliable financial search must not only provide the correct answer, but also ensure authoritative sources, accurate timing, and consistent conventions, and make key data and conclusions reviewable.
Ant Group, together with CICC, developed FinFIRST (Financial Information Retrieval, Sourcing and Traceability). This evaluation benchmark was designed with deep involvement from more than 50 financial professionals, systematically assessing from three dimensions—results, process, and evidence—whether answers are accurate and complete, whether research conventions and calculations are reasonable, and whether key data and conclusions can be traced back to authoritative primary sources.FinFIRST will be open-sourced soonproviding a reliable benchmark for financial search research and evaluation.
Evaluation results show that Ling-3.0-flash-Fin's overall score has surpassed some large flagship models,with outstanding competitiveness compared with models of the same size; in particular, it excels at prioritizing official, primary, and highly credible sources.
Demo: Google's monthly Token processing volume across all platforms
🎬 Video demo (seethe original WeChat article ）
Demo note: This demo only visualizes the processing trajectory of the current example task and does not represent the model's full capabilities, all task workflows, or the final product interface; the demo content is for reference only and does not constitute any investment advice or basis for decision-making
This case requires the model to sort out the four first-time disclosures of Google's monthly Token processing volume across all platforms and calculate the increment at each stage.
Through 23 tool calls, Ling-3.0-flash-Fin ultimately focused on Google I/O, Alphabet earnings calls, Google Cloud announcements, SEC filings, and the CEO blog, constructing a continuous data series following "broad search for authoritative sources—disambiguating conventions—locking in first disclosures—unifying data—interval calculation," and accurately reconstructing the four disclosures and the average monthly net increment for adjacent intervals.
03 Research reasoning: from data to judgment, forming reviewable research conclusions
Real financial research rarely derives answers directly from a single piece of information. Models need to synthesize information from different documents, data conventions, and time points, clarify facts and assumptions, and gradually build a complete and rigorous evidence chain through analysis and computation.
Finance Agent Bench integrates information retrieval, announcement analysis, financial data verification, and multi-step calculation into a single task. The results show that Ling-3.0-flash-Fin maintains competitive financial research performance under a smaller activated-parameter architecture, demonstrating that training enhances its capabilities in announcement analysis, information retrieval, and research tasks.
Demo: Spire business reclassification and earnings restructuring
🎬 Video demo (seethe original WeChat article ）
Demo note: This demo only visualizes the processing trajectory of the current example task and does not represent the model's full capabilities, all task workflows, or the final product interface; the demo content is for reference only and does not constitute any investment advice or basis for decision-making
In the Spire case, Ling-3.0-flash-Fin made 55 tool calls to verify the 10-K, earnings announcements, and 8-K item by item, identified the change in reporting basis after Storage was reclassified as a discontinued operation, re-split the original Midstream earnings, and cross-reconciled along the three paths of Adjusted, Other, and GAAP. The model ultimately confirmed that the historical earnings of continuing operations were $17.8 million and made clear that this figure cannot be used directly as a baseline for subsequent forecasts.
04 Valuation modeling: understanding the complex reconciliation of financial models and delivering automatic model updates
During earnings season, analysts often need to update the financial models of dozens of companies immediately after companies disclose results, organize analysis charts, and prepare research reports. Cross-document data extraction, actual-value overwriting, switching formulas between historical and forecast periods, financial reconciliation, anomaly troubleshooting, and visualization creation are all interlinked, which not only involves a heavy workload but also makes errors due to reporting basis or formula issues likely.
SpreadsheetBench V1 mainly tests cell, formula, and formatting operations in real workbooks; SpreadsheetBench V2 extends the tasks to model generation, debugging, visualization, and complex multi-sheet workflows. Relevant evaluations show that Ling-3.0-flash-Fin already has a solid foundation in structured spreadsheet operations and financial model understanding, and will continue to strengthen the stability of complex cross-sheet dependencies, anomaly recovery, and end-to-end delivery.
Demo: Google 2026 Q2 financial model update
🎬 Video demo (seethe original WeChat article ）
Demo note: This demo only visualizes the processing trajectory of the current example task and does not represent the model's full capabilities, all task workflows, or the final product interface; the demo content is for reference only and does not constitute any investment advice or basis for decision-making
In this case, Ling-3.0-flash-Fin read the company's financial report and an Excel model containing 7 worksheets and more than 5,000 formulas, identified the correspondence between actual values and forecast values, updated the 2026 Q2E forecast data to the actually disclosed 2026 Q2 data, and simultaneously handled the relevant formulas, cross-sheet dependencies, and Summary charts.
The model focused on completing three types of operations:
- matching actual values with forecast values and switching the corresponding formulas;
- restoring single-quarter data and synchronizing cross-sheet dependencies;
- checking for circular references, balancing relationships, and formula conventions, and updating the Summary sheet and analysis charts.
The final workbook retains its original structure, formula relationships, and editability.
05 Research report writing: letting data and charts jointly support research conclusions
Research report writing requires the model to identify core propositions from high-density materials, connect facts, data, and judgments, and form clear, reviewable research conclusions through appropriate charts and narrative structures.
Demo: From research materials to professional expression
🎬 Video demo (seethe original WeChat article ）
Demo note: This demo only visualizes the processing trajectory of the current example task and does not represent the model's full capabilities, all task workflows, or the final product interface; the demo content is for reference only and does not constitute any investment advice or basis for decision-making
In the following two cases, Ling-3.0-flash-Fin organized scattered materials into an evidence chain of "fact—explanation—judgment—chart":
- Micron: restoring the economic meaning of contract information. The model placed the strategic customer agreement and the approximately $100 billion of remaining performance obligations back into the context of a multi-year contract, distinguished future performance obligations from current-period revenue, and preserved judgment boundaries such as contract execution, price adjustments, and changes in supply and demand.
- NVIDIA: identifying sources of marginal growth from a new disclosure basis. Hyperscale and ACIE businesses are similar in scale, with quarter-over-quarter growth rates of 12% and 31%, respectively. Based on this, the model distilled that the marginal growth drivers may be spreading to AI cloud, sovereign AI, and enterprise and industrial customers, and mapped research judgments one by one to official disclosures through side-by-side charts.
In the existing cases, Ling-3.0-flash-Fin can organize data, analytical text, and charts around core propositions, making each piece of evidence serve the research conclusion. In terms of visual expression, the model selects chart forms and establishes information hierarchies based on content relationships, generating research materials that are convenient for subsequent editing, roadshow communication, and professional review.
06 A complete case: how four capabilities converge into the same task chain
To observe how these capabilities work together, we selected a Planet Fitness LBO (leveraged buyout) case originating from the APEX-Agents investment banking task scenario and reproduced it in the current tool environment. The task required the model to adjust future operating expense assumptions, calculate investment returns under different scenarios, and complete a sensitivity analysis table in the original workbook.
🎬 Video demo (seethe original WeChat article ）
Demo note: This demo only visualizes the processing trajectory of the current example task and does not represent the model's full capabilities, all task workflows, or the final product interface; the demo content is for reference only and does not constitute any investment advice or basis for decision-making
Ling-3.0-flash-Fin traced formula dependencies along operating expenses, EBITDA, free cash flow, debt repayment, and IRR; when the spreadsheet tool could not correctly refresh some formulas, the model switched to Python to rebuild the key calculation chain and cross-verified it against benchmark results, ultimately completing scenario return calculations, sensitivity analysis, and workbook updates while preserving file editability and key calculation bases.
07 Limitations and future plans
The complexity of the financial field determines that building professional capabilities cannot be achieved overnight.
As the first finance-enhanced version, Ling-3.0-flash-Fin has initially acquired the core capabilities required for investment research, but moving from individual capabilities to stable, complete workflows still requires continuous testing and improvement in real-world use. When key assumptions, valuation results, and investment conclusions are involved, model outputs still need to be reviewed by professionals and do not constitute any investment advice or basis for decision-making.
In the next stage, Ling will continue to improve task data, process verification, and result traceability, and enhance the coordination of each link; at the same time, it will explore finance enhancement on a larger-scale model foundation to further improve complex reasoning and long-chain task capabilities, and gradually expand to financial scenarios beyond investment research.
08 Promoting long-term evolution through open co-construction
The long-term development of financial intelligence capabilities cannot be achieved without the joint participation of financial institutions, industry experts, professional users, and developers. Ant Ling will work closely with industry partners to continuously improve the financial data and evaluation system, and promote the validation and application of model capabilities in real business scenarios.
Ling-3.0-flash-Fin will continue the open approach of the Ling series, building on an open model foundation, and will gradually open up the model, tools, and evaluations, providing financial institutions, research teams, and developers with flexible deployment and secondary development opportunities.
Ling will continue to be rooted in real financial workflows, enabling professional financial capabilities to serve a broader range of needs.
Open Intelligence for the Real World.
09 Access Now
For financial practitioners and developers, Ling-3.0-flash-Fin will offer a one-month limited-time free API access on OpenRouter, making it easy for users to quickly experience and validate the model's capabilities in scenarios such as financial research, data analysis, and application development.
OpenRouter access address:
https://openrouter.ai/inclusionai/ling-3.0-flash-fin:free#providers
The Ling-3.0-flash-Fin model weights will be officially open-sourced next week. Stay tuned.
Model ReleaseFinancePrevious postOpen Models × Open Evaluation: Ling-3.0-flash-Fin and FinFIRST Officially Open-SourcedNext postLing-3.0 tiny & flash Base Models Officially Open-Sourced