10 月 28 日，蚂蚁集团旗下百灵大模型正式开源 Ling-3.1-flash-1019。新版本在 Ling-3.1-flash 基础上进一步训练优化，在智能指数 Artificial Analysis Intelligence Index v4.3.2 中获得 44 分，较上一版本提升 3 分。与此同时，百灵同步开源其基础模型 Ling-3.1-flash-base，为长程任务所需的知识、推理和计算效率打好基础。
据了解，Ling-3.1-flash-1019 版本面向 Personal Agent 场景，进一步提升了模型处理复杂工作流和多步骤任务的表现。从测评具体来看，新版本在复杂知识工作 AA-Briefcase 中的得分从约 1400 Elo 提升至 1524 Elo，GDPval-AA 从约 1618 Elo 提升至 1644 Elo；流程自动化 AutomationBench-AA 得分从 62%提升至 68%；科学问题求解 CritPt 得分从 18%提升至 26%，SciCode 得分从 54%提升至 57%。
记者注意到，百灵本次模型迭代开源与灵光 2.0 息息相关。就在 10 月 26 日，灵光 2.0 升级为「能管事的个人 AI 助理」，依托跨会话记忆、长程任务调度与独立云主机等能力，灵光可以在用户授权范围内更主动地跟进任务、持续管理生活中的复杂事务，即使用户退出对话，相关任务也能继续运行。这背后对模型能力提出了新要求。
从回答问题到持续办事，模型需要哪些新能力？
与即时问答不同，个人智能体需要围绕一个目标持续工作，这既需要产品具备记忆和任务调度机制，也需要底层模型能够理解目标、使用工具、处理新信息并推进任务。
灵光 2.0，呈现了这类能力在个人生活中的应用。 例如，规划一趟旅行，不只是生成一份行程攻略，还需要结合出发时间跟进天气、预约和行前准备；面对双 11 购物需求，则要按照预算、规格和个人偏好筛选商品、整理清单，并持续关注价格变化，在符合约定条件时提醒用户。
围绕这些需求，百灵持续通过模型训练和应用反馈，探索 Personal Agent 所需的能力。Ling-3.1-flash-base 在设计初期就将这些需求纳入考虑，模型采用混合线性 MoE 架构，总参数量约 560B，每个 Token 激活约 25B 参数，在模型容量、长序列处理和单步计算开销之间取得平衡，为后续面向真实任务的训练提供基础。
在 8 个领域共 47 项 benchmark 测试中，Ling-3.1-flash-base 在 32 项上超越了 DeepSeek-V4-Pro-Base 和 MiMo-V2.5-Pro-Base，取得最高分。其中，知识理解、逻辑推理、代码、数学与长上下文评测，对应着个人助理处理复杂事务所需的多项基础能力，成为处理复杂任务的重要起点。据了解，Ling-3.1-flash-base 以 BF16 基础权重、MIT 许可证向开发者开放，开发者可以基于自身的知识、流程和工具开展继续预训练、领域适配及后训练。
在基础能力之上，Ling-3.1-flash 及后续的 1019 版本通过后训练，持续强化推理、工具使用和多步骤执行能力，持续提升模型在实际工作流中的表现。记者注意到，百灵从 Base 到后训练模型，研发重点逐步延伸到如何遵循目标、使用工具和推进任务。
据了解，围绕个人事务中的多模态信息处理需求，百灵基于 Ling-3.1-flash 训练了视觉语言模型 Ling-3.1-flash-VL，在既有推理、规划和工具使用能力的基础上增加视觉理解，目前已先行应用于灵光，后续将对外发布。
从真实任务中来，到真实应用中去
目前，百灵正逐步完善 Personal Agent 所需的模型能力。模型层面的能力演进，也正在应用中得到进一步验证和探索。据了解， Ling-3.1-flash 已基于灵光、阿福医生版等应用场景的需求与反馈开展对齐训练，持续提升模型在专业表达和沟通等方面的能力。
“从真实任务中来，到真实应用中去。”百灵表示，希望通过模型研发与应用实践的持续循环，提升模型理解目标、保持任务状态和应对变化的能力，并且将持续通过开源与开发者共同探索更多应用可能。
记者注意到，不同于单纯追求模型参数规模，百灵更注重让基础模型能力进入真实流程，完成真实任务。今年下半年以来，百灵围绕开放开源和真实场景加速模型迭代和发展。
此前，百灵基于新一代基座模型 Ling-3.0-flash，分别推出金融增强模型 Ling-3.0-flash-Fin 和医疗增强模型 Ling-3.0-flash-Sante，将基础模型能力进一步带入真实专业场景。 其中，Ling-3.0-flash-Fin（总参数 124B、激活 5.1B）让 AI 进入真实投研任务流，打通从找资料、做分析到形成研究成果的完整任务链，在高难度投研任务中比肩 3T 级前沿模型，金融 Agent 综合表现达到开源领先；Ling-3.0-flash-Sante 则将“高智效比”带入医疗场景，以 124B 参数规模在 1.6T 以下模型中取得最优表现，整体性能与多款旗舰模型相当。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On 28 October, the official open source of Ling-31-flash-1019 was LLM under the ants Group flag. The new version is based on Ling-3.1-flash and optimizes further, with 44 points in the Smart Analysis Inteligence Index v.4.3.2, up 3 points from the previous version. At the same time, Ling-3.1-flash-base, the base model of the Centrifugal Synchronizing Source, provides the basis for the knowledge, reasoning and computational efficiency required for a long-range mission.
It is understood that the Ling-31-flash-1019 version is oriented towards the Personal Agent scene, further enhancing the performance of the model in dealing with complex workflow and multistep tasks. Specifically, the new version of the AA-Briefcase score rose from about 1400 Elo to 1524 Elo, GDPval-AAA from about 1618 Elo to 1644 Elo, process automation Bench-AA from 62 per cent to 68 per cent, scientific resolution from 18 per cent to 26 per cent, and Cisco from 54 per cent to 57 per cent.
Journalists have noticed that the dichotomy of this model is closely related to the light 2.0. Just on October 26th, the light 2.0 was upgraded to an "AI Assistant who can run the business", based on such capabilities as inter-chamber memory, long-range tasking and independent cloud hosting, so that it can more proactively follow up on tasks within the user’s delegated authority, and continue to manage complex matters of life, even if the user withdraws from dialogue. There are new requirements behind this for modelling capacity.
What new capabilities do models require from answering questions to working on an ongoing basis?
Unlike instant questions and answers, an individual AI agent needs to work on an ongoing basis around a goal that requires both a product with memory and task control mechanisms, as well as a bottom model to understand the objective, use tools, process new information and advance the task.
The light 2.0 presents the application of these abilities in personal life. For example, a trip is planned, not only to generate a trip plan, but also to follow the weather, appointment and pre-departure in conjunction with the time of departure; in the face of double 11 shopping needs, goods are screened, lists are organized according to budget, specifications and personal preferences, and constant attention is paid to price changes and users are reminded when they meet the agreed conditions.
In the context of these needs, the Bainimarama continuously explores the capabilities needed for the Personal Agent through model training and application feedback. Ling-3.1-flash-base took these needs into account at the early stages of the design, using a hybrid linear MoE structure, total parameter count about 560B, each Token activates approximately 25B parameters, balancing model capacity, long-sequence processing and one-step costing, providing a basis for subsequent training for real task.
In a total of 47 benchmark tests in 8 areas, Ling-31.1-flash-base exceeded DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base in 32 points to achieve the highest score. Among them, knowledge understanding, logical reasoning, code, mathematics and long context assessments correspond to the many basic competencies required for personal assistants to handle complex matters and constitute an important starting point for complex tasks. Ling-31-flash-base is known to be open to developers with MF16 base weights and MIT licences, who can continue pre-, area-appropriate and post-training based on their own knowledge, processes and tools.
On the basis of basic capabilities, Ling-31-flash and subsequent 1019 versions have been trained to continuously enhance reasoning, tool use and multi-step implementation capacity and to continuously enhance the model ' s performance in actual workflow. Journalists noted that the focus of research and development from the Base to the post-training model has gradually been extended to how to follow objectives, use tools and advance tasks.
It is understood that Ling-31-flash-VL, a visual language model based on Ling-31-flash, has been trained around the need for the processing of multi-modular information in personal affairs, which builds on the existing ability to reason, plan and use tools and is now being applied first and then to external publish.
From real task to real application.
At present, the Ultimate is gradually refining the modelling capabilities required for the Personal Aget. The evolution of capabilities at the model level is also being further validated and explored in applications. Ling-3.1-flash is known to have conducted matching training based on the needs and feedback of applications such as the Shining Light, Dr. Afu, etc., to continuously upgrade the model ' s capabilities in areas such as professional expression and communication.
"from real task to real application." Benevolence expressed the hope that through the continuous cycle of model development and application practices, models would be upgraded to understand objectives, maintain mission status and respond to changes, and that more applications would be explored on an ongoing basis with developers through open sources.
Journalists noted that, unlike simply pursuing the size of the model parameters, the Bainimarama focused more on bringing basic model capabilities into the real process and completing real task. Since the second half of this year, the Bainimarama has evolved over time around open-source and real scenario acceleration models.
Prior to this, Bain Ling-3.0-flash, based on a new generation of base models, introduced the financial enhancement model Ling-3.0-flash-Fin and the medical enhancement model Ling-3.0-flash-Sante, respectively, to bring basic model capabilities further into the real professional scene. Among them, Ling-3.0-flash-Fin (total parameter 124B, activation 5.1B) allows AI to enter the real investment task stream and to access the full chain of tasks from information search, analysis to the production of research results, and to match the 3T level front-line model to the open source in the high-probability study, where the financial Agent combined performance leads to the open source; Ling-3.0-flash-Sante brings "high-wissper" to the medical scene with the best performance at the scale of the 124B parameter in the 1.6T model, with overall performance comparable to that of multi-shipship.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.