On October 28, the Ling-3.1-flash-1019 large model under Ant Group was officially made open source. The new version is further trained and optimized based on Ling-3.1-flash. It scored 44 points in the Artificial Analysis Intelligence Index v4.3.2, an increase of 3 points compared to the previous version. Meanwhile, Bailing also made its base model Ling-3.1-flash-base open source, laying the foundation for knowledge, reasoning, and computational efficiency required for long-distance tasks.
It is understood that the Ling-3.1-flash-1019 version is designed for the Personal Agent scenario, and further improves the model’s performance in handling complex workflows and multi-step tasks. Specifically, in the complex knowledge work AA-Briefcase assessment, the score increased from approximately 1400Elo to 1524Elo, while GDPval-AA rose from approximately 1618Elo to 1644Elo. The AutomationBench-AA score improved from 62% to 68%, and the CritPt score for scientific problem solving increased from 18% to 26%, with SciCode scoring rising from 54% to 57%.
The reporter noticed that this iteration of the Bai Ling model’s open-source release is closely related to Lingguang 2.0. On October 26, Lingguang 2.0 was upgraded to a “capable personal AI assistant”. With capabilities such as cross-session memory, long-term task scheduling, and independent cloud hosts, Lingguang can more proactively follow up on tasks and continuously manage complex matters in daily life within the scope of user authorization. Even if the user exits the conversation, relevant tasks can continue to run. This has set new requirements for the model’s capabilities.
What new capabilities does the model need to go from answering questions to continuing tasks?
Unlike instant Q&A, personal intelligent agents need to work continuously around a goal. This requires the product to have memory and task scheduling mechanisms, as well as the underlying model to understand the goal, use tools, process new information, and advance tasks.
Lingguang 2.0 presents the application of such capabilities in personal life. For example, when planning a trip, it isn’t just about generating a itinerary; it also requires tracking weather conditions, making reservations, and preparing for the journey based on the departure time. When dealing with the Double 11 shopping demand, users need to filter products according to budget, specifications, and personal preferences, create a list, and continuously monitor price changes, reminding users when the agreed conditions are met.
Around these requirements, Ling-3.1-flash-base continues to explore the capabilities necessary for a Personal Agent through model training and application feedback. These requirements were taken into consideration from the initial design phase of Ling-3.1-flash-base. The model adopts a hybrid linear MoE architecture with a total parameter count of approximately 560B, and about 25B parameters are activated for each Token. It achieves a balance between model capacity, long-sequence processing, and single-step computation overhead, providing a foundation for subsequent training focused on real-world tasks.
In 47 benchmark tests across 8 domains, Ling-3.1-flash-base surpassed DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base on 32 tests, achieving the highest score. Specifically, knowledge comprehension, logical reasoning, code, mathematics, and long context evaluation represent several fundamental abilities required for a personal assistant to handle complex tasks, making them important starting points for processing complex missions. It is understood that Ling-3.1-flash-base is open to developers with a BF16 base weight and the MIT license. Developers can conduct continued pre-training, domain adaptation, and post-training based on their own knowledge, processes, and tools.
Above the basic capabilities, after passing the training with Ling-3.1-flash and the subsequent 1019 version, the reasoning, tool usage, and multi-step execution abilities are continuously enhanced, steadily improving the model’s performance in actual workflows. The reporter noticed that from the Base model to the post-training model, the research focus gradually shifted to how to follow objectives, use tools, and advance tasks.
It is understood that in response to the multi-modal information processing needs in personal affairs, Bailing has trained a visual language model called Ling-3.1-flash-VL using Ling-3.1-flash. On top of its existing reasoning, planning, and tool usage capabilities, visual understanding has been added. This model has already been applied in Lingguang, and it will be made available to the public later.
From real tasks to real applications
Currently, Ling-3.1-flash is gradually improving the model capabilities required by Personal Agent. The evolution of model capabilities is also being further verified and explored through applications. It is understood that Ling-3.1-flash has undergone alignment training based on the needs and feedback from various application scenarios such as Lingguang and Dr. Afu, continuously enhancing the model’s abilities in professional expression and communication.
“From real tasks to real applications.” Bailing said that he hopes through the continuous cycle of model development and application practice, the ability to understand the goals of models, maintain task status, and respond to changes will be improved. Moreover, more application possibilities will continue to be explored through open source and collaboration with developers.
The reporter noticed that, unlike simply pursuing the scale of model parameters, Bailing focuses more on bringing the capabilities of the basic model into real-world processes and completing actual tasks. Since the second half of this year, Bailing has accelerated the iteration and development of models around open source and real-scenario applications.
Previously, Bailing introduced the financial enhancement model Ling-3.0-flash-Fin and the medical enhancement model Ling-3.0-flash-Sante based on the new generation base model Ling-3.0-flash, further bringing the capabilities of the basic model into real professional scenarios. Among them, Ling-3.0-flash-Fin (with a total parameter size of 124B and 5.1B activated) enables AI to engage in real investment research tasks, completing the entire task chain from data retrieval, analysis, to research outcomes. In high-difficulty investment research tasks, it competes with 3T-level cutting-edge models, and its overall performance as a financial Agent exceeds that of open-source models; Ling-3.0-flash-Sante brings “high intelligence efficiency” into the medical domain, achieving optimal performance in models below 1.6T with a parameter size of 124B, and its overall performance is comparable to several flagship models.