July 28, atWorld Artificial Intelligence ConferenceOn, Ant Group's technology subsidiaryAnt Count FamilyOfficial ReleaseThe Great Model of Financial Reasoning Agentar-Fin-R1.
It is reported that Agentar-Fin-R1 is not only a new model, but also a "professional hub" tailor-made for financial scenarios such as banking, securities, insurance, etc., focusing on "reliable, controllable and optimizable" .
Specifically, Agentar-Fin-R1 is based on Qwen3 and outperforms Deepseek-R1 on benchmarks such as FinEval1.0, FinanceIQ, and other authoritative financial big models, as well as financial big models of the same size as open source generic big models.
performance performance.Agentar-Fin-R1's general-purpose ability, such as the 32B version, scored 93.80 on MATH and 68.18 on GPQA; its financial ability is even more outstanding, with the 32B model comprehensively outperforming high-parameter models such as DeepSeek-R1 and GPT-o1 on the two major financial benchmarks, FinEval1.0 and FinanceIQ. Generalized models such as DeepSeek-R1, GPT-o1 and other high-parameter models.
The R&D team has also constructed an extremely comprehensive and professional financial data corpus in the industry. A financial task system covering all scenarios of banks, securities, insurance, funds, trusts, etc., including 6 major categories and 66 sub-scenarios, constitutes the most systematic and real financial data set in the industry. What's more special is that "principle-based synthetic data" is also introduced in the training, so that the model naturally complies with the red line of financial regulation.
Currently, Agentar-Fin-R1 is available in 32B and 8B parameterized versions, in addition to the MOE architectural model based on the Beringia model, and 14B and 72B non-inferential versions.