Ant Group open-sourced Ling-3.0-flash-Fin on September 9, 2026, at the Inclusion Conference on the Bund in Shanghai. The model is the first finance-tuned checkpoint in the Ling 3.0 line. Weights on Hugging Face list 124 billion total parameters, 5.1 billion active per token, and a 256,000-token context window, built as a Mixture-of-Experts stack that Ant says should hold large-model range at small-model serving cost.

The Business Wire release said the lab worked with financial institutions and domain experts so the model can retrieve from official sources, stitch multi-document evidence, update Excel valuation models, and draft research reports with charts. Ant also put the checkpoint on ModelScope, OpenRouter, and Vercel, and said teams can run it privately and attach search, Python, databases, and spreadsheets.

Alongside the weights, Ant released FinFIRST, a benchmark written with the investment banking team at China International Capital Corporation. FinFIRST V1 has 123 expert tasks, 701 atomic criteria, and 12,300 rubric points. The point of the rubric is to score the research process, not only the final number.

The model card says Ling-3.0-flash-Fin was continued from Ling-3.0-flash on financial data, that thinking mode is on by default, and that the checkpoint is competitive on FinSearchComp, FinCRAFT, Finance Agent, APEX-Agents, SpreadsheetBench, and τ³-Banking. Ant is explicit that valuation outputs need professional review and are not investment advice. The same family now includes Ling-3.0-tiny, Ling-3.0-flash-VL, and Ling-3.0-flash-Santé for local, multimodal, and healthcare work.

Decoded Take

A finance model that can touch a live spreadsheet is only useful if the citations stay attached when the file leaves the chat. Ant is pairing weights with a CICC-authored rubric, which is a smarter open-source move than another leaderboard screenshot, because buy-side reviewers will argue about source selection long before they argue about parameter counts. The risk is familiar: a 5.1 billion active MoE can look cheap in a demo and still hallucinate a footnote once the filing language gets ugly. Watch whether independent desks reproduce FinFIRST scores on their own binders, whether the card’s “needs professional review” line survives the first public valuation error, and whether the Santé and VL siblings get the same process-level benches or only marketing labels.