Dynamic Beating AI News: Antbird has officially open-sourced the financial model Ling-3.0-flash-Fin. The model has a total of 124 billion parameters, 5.1 billion activation parameters, and supports a context of 256K. The officially released BF16 weights are around 255GB, using the MIT license, and can be deployed independently with vLLM and SGLang.
Also open-sourced is the financial search benchmark FinFIRST. Developed by Ant Group with professional support from the CICC Investment Banking Team, over 50 financial professionals were involved in question setting and validation. In the end, only 9.78% of the candidate questions passed the screening, leaving 123 questions. Among them, 61.8% require calculations, and 32.5% require simultaneous referencing of multiple sources.
Each question will be broken down into multiple scoring points to check whether the AI has identified the data correctly, used the right sources, validated the timing and statistical methods, and if the calculation results are accurate. The entire set of data consists of 701 detailed scoring points. Now, the questions, reference answers, and scoring criteria are all publicly available under the Apache 2.0 license.

