On October 28, Ant Group's Bailian (Bailong) LLM was officially open-sourced as Ling-3.1-flash-1019. The new version was further trained and optimized on top of Ling-3.1-flash, scoring 44 points on the Artificial Analysis Intelligence Index v4.3.2, an improvement of 3 points over the previous version. At the same time, Bailian also open-sourced its base model Ling-3.1-flash-base, laying the foundation for the knowledge, reasoning, and computational efficiency required for long-horizon tasks.
According to reports, the Ling-3.1-flash-1019 version targets Personal Agent scenarios and further improves the model's performance in handling complex workflows and multi-step tasks. Specifically, in evaluations, the new version's score in complex knowledge work AA-Briefcase rose from about 1400 Elo to 1524 Elo, and GDPval-AA rose from about 1618 Elo to 1644 Elo; in workflow automation, the AutomationBench-AA score improved from 62% to 68%; in scientific problem solving, the CritPt score improved from 18% to 26%, and the SciCode score improved from 54% to 57%.
The reporter noted that Bailing's open-sourcing of this model iteration is closely related to Lingguang 2.0. On October 27, Lingguang 2.0 was upgraded to a "personal AI assistant that can manage affairs." Relying on capabilities such as cross-session memory, long-term task scheduling, and an independent cloud host, Lingguang can more proactively follow up on tasks and continuously manage complex affairs in users' lives within the scope of user authorization; even if the user exits the conversation, related tasks can continue to run. This places new demands on model capabilities behind the scenes.
What new capabilities does a model need to move from answering questions to continuously getting things done?
Unlike instant question answering, personal agents need to work continuously toward a goal, which requires the product to have memory and task scheduling mechanisms, and also requires the underlying model to be able to understand goals, use tools, process new information, and advance tasks.
Lingguang 2.0 demonstrates how these capabilities apply to personal life. For example, planning a trip involves not just generating an itinerary, but also tracking weather, reservations and pre-trip preparations based on the departure time; facing the Double 11 shopping demand, it must filter products and compile lists according to budget, specifications and personal preferences, continuously monitor price changes, and remind users when agreed conditions are met.
Around these needs, Bailing has continuously explored the capabilities required of a Personal Agent through model training and application feedback. Ling-3.1-flash-base took these needs into account from the early design stage. The model adopts a hybrid linear MoE architecture with approximately 560B total parameters and approximately 25B parameters activated per Token, striking a balance between model capacity, long-sequence processing and per-step computation cost, laying a foundation for subsequent training oriented toward real tasks.
In benchmark tests spanning 47 items across 8 domains, Ling-3.1-flash-base achieved the highest scores on 32 items, surpassing DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base. Among them, knowledge understanding, logical reasoning, code, math and long-context evaluations correspond to many of the foundational capabilities a personal assistant needs to handle complex affairs, serving as an important starting point for handling complex tasks. It is understood that Ling-3.1-flash-base is made available to developers with BF16 base weights under the MIT license, and developers can conduct continued pre-training, domain adaptation and post-training based on their own knowledge, processes and tools.
On top of the foundational capabilities, Ling-3.1-flash and the subsequent 1019 version, through post-training, continuously strengthen reasoning, tool use and multi-step execution capabilities, steadily improving the model's performance in real workflows. The reporter noted that from Base to post-trained models, Bailing's R&D focus has gradually extended to how to follow goals, use tools and advance tasks.
It is understood that, addressing the multimodal information processing needs in personal affairs, Bailing trained the vision-language model Ling-3.1-flash-VL based on Ling-3.1-flash, adding visual understanding on top of its existing reasoning, planning and tool-use capabilities. It has already been deployed first in Lingguang and will be released externally later.
From real tasks, to real applications
Currently, Bailing is progressively refining the model capabilities required of a Personal Agent. The evolution of capabilities at the model level is also being further validated and explored in applications. It is understood that Ling-3.1-flash has already carried out alignment training based on the needs and feedback from application scenarios such as Lingguang and the Afu Doctor Edition, continuously improving the model's capabilities in professional expression and communication.
"From real tasks, to real applications." Bailing stated that it hopes to improve the model's ability to understand goals, maintain task state and cope with changes through a continuous loop of model R&D and application practice, and will continue to explore more application possibilities together with developers through open source.
The reporter noted that, unlike merely pursuing model parameter scale, Bailing focuses more on bringing base model capabilities into real workflows to complete real tasks. Since the second half of this year, Bailing has accelerated model iteration and development around open source and real-world scenarios.
Previously, based on the new-generation base model Ling-3.0-flash, Bailing launched the finance-enhanced model Ling-3.0-flash-Fin and the healthcare-enhanced model Ling-3.0-flash-Sante, further bringing base model capabilities into real professional scenarios. Among them, Ling-3.0-flash-Fin (124B total parameters, 5.1B activated) brings AI into real investment research task flows, connecting the complete task chain from finding materials, conducting analysis to forming research results; in high-difficulty investment research tasks it rivals 3T-class frontier models, and its overall financial Agent performance reaches an open-source leading level. Ling-3.0-flash-Sante brings a "high intelligence-efficiency ratio" into the healthcare scenario, achieving the best performance among models under 1.6T at a 124B parameter scale, with overall performance comparable to several flagship models.
This article was provided by Ant Group; QbitAI republished it with authorization, and the views belong to the original author.