A fund analyst spends the morning rewriting the same valuation model in three tools, copying numbers between a spreadsheet, a research terminal, and a chat window. The bottleneck is rarely raw intelligence — it is plumbing, audit trails, and trust. Ant Group's latest release is a bet that the next wave of useful AI in finance is not a bigger brain, but a cheaper, more traceable one.
Key takeaways
- Ant open-sourced Ling-3.0-flash-Fin at the 2026 Inclusion·Conference on the Bund (外滩大会), a finance-tuned model built on a Mixture-of-Experts (混合专家) architecture with 124B total parameters but only 5.1B active per token.
- It targets four workflow capabilities — retrieval with traceable sources, research reasoning, valuation modeling inside Excel, and report generation — co-developed with financial institutions including the investment banking team at CICC (中金公司).
- Ant also released the wider Ling 3.0 family: a production agent model, a 7.9B/1.3B "tiny" model that runs fully offline, a vision-language variant, and a healthcare-tuned model.
- The weights are on Hugging Face and ModelScope; the API ships on OpenRouter and Vercel, with an open release rather than a walled garden.
- The thesis: in finance, trustworthy, end-to-end-traceable outputs matter more than leaderboard points, and token efficiency translates directly into cost.
Why a finance model, and why now
Generic large models (大模型) are good at answering questions and terrible at the unglamorous work of finance: pulling the right figure from an authoritative source, keeping a calculation consistent across 40 tabs, and showing the reader exactly where every number came from. Ant's Bailian (百灵) team is arguing that a domain-shaped model, trained against real workflows, beats a generalist on the metrics that actually matter to a bank or a buy-side desk.
The release lands at a moment when Chinese internet giants are pulling AI development in-house. Ant is not selling a chatbot; it is selling a component that can be dropped into compliance-heavy pipelines where every output must be explained to a regulator.
What "124B total, 5.1B active" actually means
Ling-3.0-flash-Fin uses a sparse Mixture-of-Experts design. Of its 124 billion parameters, only about 5.1 billion switch on for any single token. That keeps inference cheap and deployment light while preserving the knowledge capacity of a far larger model. For a firm running millions of queries a day, the math is simple: a smaller active footprint means a smaller bill and lower latency, without surrendering the breadth a financial task often demands.
Ant has been consistent about this efficiency posture. The earlier Ling-2.6-flash API was listed at US$0.1 per million input tokens and US$0.3 per million output tokens — prices set to make high-volume agent use economically sane.
The four capabilities Ant is selling
The finance model is organized around four concrete jobs rather than abstract "intelligence":
- Information retrieval that prioritizes authoritative, official sources and keeps data consistent and end-to-end traceable.
- Research reasoning that synthesizes messy, heterogeneous data into a verifiable chain of evidence.
- Valuation modeling that understands complex Excel linkages, supports automated updates, and keeps files fully editable.
- Report generation that fuses facts, calculations, and charts into readable professional output.
These are the tasks a first-year analyst actually does, repeated thousands of times. Framing the model around them is a clear signal that Ant is aiming at back-office productivity, not consumer chatter.
FinFIRST: an open benchmark as a trojan horse
Alongside the weights, Ant open-sourced FinFIRST, a benchmark for financial search agents. It contains 123 expert-authored tasks, 701 atomic criteria, and 12,300 rubric points, built with professional support from CICC's investment banking team. Crucially, it scores the full research process rather than just matching the final answer — a direct attack on the habit of gaming leaderboards with verbose outputs.
Releasing the benchmark is a smart move: it gives third parties a yardstick and quietly sets the terms of the debate around what "good" means in financial AI.
The rest of the Ling 3.0 family
The finance model is one member of a broader portfolio. Ling-3.0-flash is a hybrid-reasoning MoE model (124B total, 5.1B active) aimed at production-scale agents. Ling-3.0-tiny is a native hybrid-reasoning model with 7.9B total and 1.3B active parameters, designed to run entirely locally with no cloud dependency — pitched at personal knowledge management and offline tasks. Ling-3.0-flash-VL adds image and video inputs with a 1M-token context for document intelligence and medical report interpretation. Ling-3.0-flash-Santé is tuned for healthcare and life sciences, claiming top-tier results among flash-size models on benchmarks like MedXpertQA-Text and DiagnosisArena-MCQ.
According to Li An, chief scientist at BrainNet (脑机网), the Chinese open-model race is shifting from "who is biggest" to "who is cheapest per useful task," and finance is the first domain where that economics really bites.
Where this fits in China's open-model push
Over the past year, a long list of Chinese labs have moved from closed APIs to open weights. Ant's angle is different from the "release the biggest model" reflex: it is domain specialization plus efficiency plus a willingness to hand researchers the benchmark that judges it. If the strategy works, the winners in enterprise AI will not be the models with the highest Intelligence Index, but the ones that survive contact with a compliance officer.
Honest limitations
- Every parameter, benchmark, and capability claim in this article traces back to Ant Group's own announcement and accompanying materials. Independent third-party evaluation of Ling-3.0-flash-Fin is still limited, and we did not run the model ourselves.
- "Open weights" does not mean "ready for regulated use." Finance deployments face data-residency, audit, and licensing constraints that a downloadable checkpoint does not solve.
- Performance on proprietary, real-world financial data — as opposed to the curated FinFIRST set — remains unverified by outside parties.
- Pricing, availability, and model variants can change quickly; treat the numbers above as a snapshot of the launch announcement.
- The co-development with CICC describes a professional-support relationship; it should not be read as an endorsement of any specific product by that institution.
What readers can do now
- Developers: pull the weights from Hugging Face or ModelScope and run FinFIRST as a first-pass yardstick on a private finance task before trusting any output.
- Finance teams: pilot the retrieval-plus-traceability workflow on a narrow use case (e.g., earnings-note drafting) and insist every figure ships with its source citation.
- Everyone evaluating AI: start measuring cost and accuracy per useful task, not parameter count — token efficiency is becoming the real competitive axis.