Ling-3.0-Flash, the new native hybrid reasoning model from Ant Bailing, was officially launched on July 24. The model has a total of 124B parameters and a single computation activation of 5.1B. While significantly reducing its size and lowering computing costs, Ling-3.0-Flash matches or even exceeds industry-leading models with 2 to 3 times its parameter scale in core metrics such as basic reasoning, instruction following, and long text processing, demonstrating a higher "intelligence efficiency ratio" and cost-effectiveness for practical applications. It is reported that the model has been launched on OpenRouter, offering free access for a limited time of one week, after which it will be officially open-sourced.
As a key feature of this version, Ling-3.0-Flash has been deeply refined for actual agent application scenarios. The model has expanded over 10,000 real interaction training environments and optimized mechanisms for self-correction and long-term planning in complex tasks. Whether in code writing, complex task decomposition, or in-depth research across multiple sources of information, Ling-3.0-Flash demonstrates stronger autonomous closed-loop delivery capabilities, solving the problem of traditional models easily "going off track" or failing in large tasks.
According to the introduction, the key to achieving "small size, high intelligence" lies in the re-design of the model's underlying computing architecture. Ling-3.0-Flash abandoned the approach of simply increasing parameters and adopted a hybrid attention architecture from the pre-training stage, alternating KDA linear attention layers and MLA layers in a 5:1 ratio, achieving a better balance between long context efficiency and model capability.
In addition, Ling-3.0-Flash upgraded the previous Lightning Attention to KDA (Kimi Delta Attention), introducing fine-grained diagonal gating in the Delta Rule state update, enabling the model to more accurately remember key information when processing long documents and code repositories. Meanwhile, Ling-3.0-Flash further compressed the expert activation ratio per computation, reducing the expert activation ratio per token from 1/32 in the previous version to 1/64, thereby achieving a higher "efficiency leverage."
To make AI agents run faster and more stably, Bailing has also provided a supporting engineering and collaboration architecture. In terms of response speed, by introducing a hierarchical caching system at the cluster level, repeated calculations in long conversations and multi-turn interactions are avoided, reducing the first-character response latency under long input by more than 60% to 80%; in terms of task stability, the upgraded multi-agent collaboration architecture allows different agents to work together and check each other, effectively reducing the risk of errors from a single model, providing support for high-frequency online services and real business applications.
