Ant Group, via its Ant Bailing unit, has officially open-sourced the financial model Ling-3.0-flash-Fin. Big model. It carries 124 billion total parameters, with 5.1 billion activated parameters, and it supports a 256K context window. The release also ships BF16 weights at roughly 255GB, under the MIT license, so developers can deploy it with vLLM or SGLang.
And with the model, Ant Group also released the FinFIRST financial search benchmark. It was built together with the investment banking team at China International Capital Corporation (CICC). More than 50 financial professionals took part in writing and validating the questions. The filter was tough: just 9.78% of submitted questions made it through, leaving 123 final questions.
Of those 123 questions, 61.8% require computation, and 32.5% demand cross-referencing across multiple sources. Each question is broken into multiple scoring points that test whether the AI identifies the right data, cites accurate sources, checks time and statistical standards, and gets the calculations right. The full dataset contains 701 individual scoring points. And all questions, reference answers, and scoring rubrics are now public under the Apache 2.0 license.