What It Takes to Turn a General Model into a Domain Expert

A general model can discuss a domain long before it can be trusted to work in one. It can explain a financial term, summarize a filing, or describe an interface. But a professional task asks for more: the right source, the right entity, the right reporting period, a calculation that can be checked, and a result that remains useful when another person has to review it.

That gap, between knowing the language of a domain and meeting its operating standard, is where post-training matters. We are extending Ling-3.0-flash through domain continued training, task-specific post-training, tool-use optimization, expert-built evaluation, and native multimodal feedback. The aim is not to narrow a general model into a single-purpose system. It is to give it a more rigorous way to work when the task demands it.

Our recent releases make that thesis concrete. Ling-3.0-flash-Fin trains a general foundation for the linked tasks of financial research. FinFIRST makes the model’s evidence and reasoning process inspectable. Ling-3.0-flash-VL applies the same discipline to visual environments, where a model must learn from the consequence of an action rather than treating perception as the finish line.

A model that knows finance is not yet a financial model

Financial research is a useful place to start because a seemingly simple question is usually a chain of dependent checks. Why did a company’s margin change? What is the implied value under a new operating assumption? What changed between a forecast and an actual result? Each question can require a search across annual reports, earnings releases, filings, estimates, spreadsheets, and expert commentary. A number can be factually correct and still be wrong for the work: it may refer to the wrong period, currency, unit, entity, accounting definition, or version of a dataset.

Ling-3.0-flash-Fin is our first financial enhancement of Ling-3.0-flash. Built on the same sparse MoE architecture—124B total parameters, with 5.1B activated per inference step and a 256K context window—it combines high-quality financial continued training, domain post-training, and tool-use optimization. Its target is not financial Q&A. It is the chain of work around an answer: retrieve, verify, calculate, model, write, and preserve enough context that the work can be inspected.

That means teaching the model to move between the formats that researchers actually use. In a spreadsheet task, it may need to read a company report and a seven-sheet Excel model with more than 5,000 formulas; reconcile a 2026 Q2 forecast with reported actuals; update linked formulas across sheets; and refresh the summary tables and charts without breaking the model’s formula lineage. In an LBO scenario, it may need to revise operating assumptions, recalculate EBITDA, free cash flow, debt amortization, and IRR, then retain an editable output rather than merely narrating a result.

The same discipline applies to research writing. A useful draft is not a sequence of observations; it is a chain from fact to explanation to judgment to chart. The model needs to surface the source for a change in a company’s business, distinguish an existing contractual obligation from future revenue, reconcile multi-source growth drivers, and turn the work into material a researcher can still edit and challenge.

We evaluate Ling-3.0-flash-Fin across FinFIRST, FinSearchComp Verified, FinCRAFT, Finance Agent, APEX-Agents, SpreadsheetBench, and T³-Banking. Together, these cover financial retrieval, investment research, financial process execution, multi-step tool use, valuation and modeling, and spreadsheet operations—rather than treating finance as a single question-answering benchmark.

Ling-3.0-flash-Fin evaluation across financial tasks.

The difficult part is everything around the answer

A good evaluation must measure the same things that make a professional task difficult. We released FinFIRST—Financial Information Retrieval, Sourcing and Traceability—alongside Ling-3.0-flash-Fin for that reason. Developed with support from financial specialists, FinFIRST is an open benchmark for financial search agents that asks not only whether a model can find an answer, but whether it can establish why the answer should be believed.

FinFIRST benchmark overview.

The design starts from a simple observation: finance errors are often almost correct. A response can cite a real figure while using the wrong entity, date, reporting scope, currency, unit, or accounting definition. It can retrieve an authoritative source that is no longer current. It can perform a correct calculation on inputs that should never have been combined. A final-answer score does not reveal these failure modes, and a domain team cannot fix what it cannot see.

FinFIRST therefore breaks a task into atomic checks. It evaluates the correctness of the underlying information—entity, reporting period, unit, definition, and data version; the validity of citations—authority, relevance, time, and version; and the integrity of the work itself—calculation, reasoning, synthesis, and whether the conclusion is supported by evidence. Models run in a unified ReAct agent environment with Web Search, Visit, and Python. The benchmark reports Atomic, Loose Pass, and Strict Pass, so an answer that is broadly useful is distinguishable from one that passes every required check.

FinFIRST atomic scoring structure.

Its first release contains 123 expert-authored and validated tasks from more than 50 financial professionals, spanning active retrieval, financial documents, fact verification, multi-source integration, and atomized scoring. The tasks contain 701 atomic scoring items and 12,300 rubric points. They cover Chinese and English questions, multiple markets, explicit calculation, and tasks in which the model must discover and assess the evidence itself rather than being handed a direct citation. The data-construction process includes independent answering, cross-validation, rubric review, stress testing, and consistency checks; the published task-quality pass rate is 97.8%.

FinFIRST task composition and coverage.

This is the central reason to release the model and benchmark together. An open model lets teams deploy in local or private environments, govern data access, permissions, and logs, and connect their own web search, document parsing, Python, database, and spreadsheet tools. An open evaluation gives them a way to test those deployments against a standard that is closer to the work they expect the model to perform.

Vision is another way of testing the same idea

Ling-3.0-flash-VL extends the same approach to work that includes screens, documents, images, and video. The challenge is not simply to add a visual input channel. A model that can describe an image but cannot use what it sees to change its next action has not closed the loop.

Ling-3.0-flash-VL visual feedback loop.

There is a common concern that adding visual capability will trade away text performance. In our original multimodal co-training results, Ling-3.0-flash-VL improved by four points over text-only Ling-3.0-flash on the Artificial Analysis Intelligence Index v4.1.1. It also showed competitive results on image-to-webpage generation in Image-to-WebDev Arena. The important result is not that vision has become a separate feature; it is that visual information can improve the model’s representation of concepts, spatial relationships, and real-world context without turning the text foundation into a casualty.

Ling-3.0-flash-VL multimodal benchmark results.

For professional use, the relevant pattern is a visual feedback loop: observe → act → verify → correct. In front-end development, a model can read a reference image, generate code, render the page, compare the output with the target, identify the mismatch, and keep editing. In GUI work, it can carry task state across a browser, a spreadsheet, and a chart, updating its plan as the interface changes. In document intelligence, it can use images and cross-document context to surface relevant fields and relationships. In assisted medical-report reading, it can organize related indicators and flag references for review; it is not a substitute for clinical diagnosis.

The efficiency matters because these loops are iterative. Ling-3.0-flash-VL has 124B total parameters and activates 5.5B per inference step, with a native 256K context window. It supports image, text, and video input, allowing an agent to retain task history across long documents, longer videos, and multi-step execution rather than treating every observation as a new, isolated prompt.

What we mean by a domain expert

We do not mean a model that can recite more industry terminology, nor one that must be separated from general intelligence. We mean a model that has been trained against the real unit of work: it knows what evidence counts, which tools the task requires, which constraints determine a valid answer, and how to expose the reasoning and artifacts a reviewer needs.

Ling-3.0-flash-Fin shows this in an evidence-heavy research domain. FinFIRST turns the standard of proof into something developers can inspect, reproduce, and improve. Ling-3.0-flash-VL applies the same principle where the feedback arrives through a screen, a document, an image, or a video. The releases are different, but the development loop is the same: define the work with domain experts, train on the capability that work requires, equip the model to operate with tools, measure the failure modes that matter, and continue improving in the open.

That is what it takes to turn a general model into a domain expert: not a new label, but a tighter relationship between training, execution, evidence, and evaluation.