Ant Group Releases Finance-Specialized Open-Weight Model 'Ling-3.0-flash-Fin'

Ant Group has released the finance-specialized open-weight model 'Ling-3.0-flash-Fin'. This model adopts an MoE (Mixture of Experts) structure that activates only 5.1B parameters per token out of a total of 124B. It scored 23 points on the Artificial Analysis Intelligence Index, demonstrating similar performance with fewer resources compared to Ling-3.0-flash-VL (25 points), which uses 5.5B active parameters. Additionally, compared to MiniMax-M2.7 (10B active), it achieves an equivalent level of intelligence index with approximately half the active parameters.

On the finance-specialized evaluation, the Artificial Analysis Finance & Accounting Index, Ling-3.0-flash-Fin recorded the same 24 points as Ling-3.0-flash-VL. Business knowledge accuracy was 17%, higher than VL (11%), but the business knowledge hallucination rate was also higher at 33% compared to VL (19%), resulting in a non-hallucination rate of only 67%.

In the specialized knowledge task evaluation (GDPval-AA v2), it recorded 1171 Elo, lower than VL (1225) but higher than MiniMax-M2.7 (1087). In the complex business workflow evaluation (AA-Briefcase), it scored 967 Elo, slightly lower than VL (986), and trailed behind VL in analysis quality Elo (866 vs 907) and rubric pass rate (23.5% vs 24.9%). However, it showed higher scores than VL in presentation generation capability (Presentation Elo 1095 vs 1076). Output token usage averaged 67k, approximately 34% more than VL (approx. 50k) and 3.2 times that of MiniMax-M2.7 (approx. 21k).