The Zhitong Finance App learned that Ant Lairing's next-generation native hybrid inference model Ling-3.0-Flash is officially open source. LING-3.0-Flash uses an MoE architecture with 124b total parameters and 5.1b activation parameters. It retains the knowledge and capacity of the 100 billion parameter model, but only activates about 5.1 B of the parameters when generating each token, thereby reducing the computational cost of a single inference.
Developers can either directly call the model, or freely verify, deploy, and customize model capabilities, and connect LING-3.0-Flash to their own business systems, development tool chains, and agent workflows. According to reports, LING-3.0-flash provides three different landing options:
API calls allow agent applications to be launched quickly: For developers who want to quickly access and do not need to build their own inference services, Ling-3.0-Flash can be directly called through the cloud API;
Stand-alone privatization to keep data local: For enterprises and teams where data cannot be exported, the MXFP4 and INT4 versions can complete end-to-end inference on a single NVIDIA DGX Spark;
High-performance deployment, unleashing the ultimate single request speed: For high-performance services that are sensitive to single request latency, the average output rate exceeded 1100 tokens/s under the specified GPU test configuration.
In the AA Intelligence Index list, the weighted average call cost of each LING-3.0-flash task is about 0.04 US dollars, and the weighted average decoding time is about 1.4 minutes. At the same time, it enters the advantage area of “intelligence level - task cost” and “intelligence level - task time”.
LING-3.0-Flash lowers the starting threshold for privatized applications through stand-alone deployment; for real-time services that are more sensitive to single-request latency, LING-3.0-Flash further compresses generation and waiting through deep collaboration between high-performance GPUs and inference stacks, providing a lower latency operation experience for real-time agents, multi-round tool calls, and continuous task execution.