Scan the QR code to follow QbitAI

< img id=“wx_img” src=“https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png” width=“400” height=“400”>
Friends of QbitAI 2026-10-28 12:04:48 Source: QbitAI
On October 28, Ant Group’s Bailing Foundation Model officially open-sourced Ling-3.1-flash-1019.
On October 28, Ant Group’s Bailing Foundation Model officially open-sourced Ling-3.1-flash-1019. This new release builds upon Ling-3.1-flash with further training and optimization, scoring 44 on the Artificial Analysis Intelligence Index v4.3.2—a 3-point increase over the previous version. Concurrently, Bailing open-sourced its base model, Ling-3.1-flash-base, laying a solid foundation in knowledge, reasoning, and computational efficiency required for long-horizon tasks.

Engineered for Personal Agent scenarios, Ling-3.1-flash-1019 further improves performance when handling complex workflows and multi-step tasks. In benchmark evaluations, the new version saw its score on the complex knowledge task AA-Briefcase rise from ~1400 Elo to 1524 Elo, and on GDPval-AA from ~1618 Elo to 1644 Elo; process automation benchmark AutomationBench-AA improved from 62% to 68%; scientific problem-solving benchmark CritPt increased from 18% to 26%, and SciCode climbed from 54% to 57%.
This open-source model iteration is closely tied to Lingguang 2.0. Just yesterday, on October 27, Lingguang 2.0 was upgraded into “a personal AI assistant capable of managing real-world tasks.” Powered by cross-session memory, long-horizon task scheduling, and isolated cloud hosting, Lingguang can proactively track tasks and manage complex daily affairs within user-authorized boundaries. Even if the user leaves the chat, related tasks continue to execute in the background—presenting brand-new demands on underlying model capabilities.
From Answering Questions to Completing Tasks: What New Capabilities Are Needed?
Unlike instant Q&A, a personal agent must work continuously toward a specified goal. This requires not only memory and task-scheduling mechanisms at the product level, but also an underlying foundation model that can understand goals, invoke tools, process incoming information, and drive tasks forward.
Lingguang 2.0 demonstrates how these capabilities apply to daily life. For instance, planning a trip involves far more than generating an itinerary: it requires monitoring weather updates based on departure times, handling bookings, and preparing checklists. When dealing with Double 11 shopping demands, it must filter products, compile shopping lists based on budgets, specs, and personal preferences, continuously track price fluctuations, and notify the user once predefined conditions are met.
Guided by these demands, Bailing has continuously explored the essential capabilities for Personal Agents through model training and application feedback. Ling-3.1-flash-base factored these requirements into its architectural design from day one. Utilizing a hybrid linear MoE architecture with roughly 560B total parameters and ~25B activated parameters per token, it strikes an optimal balance between model capacity, long-context processing, and per-step compute costs, providing a robust base for downstream real-world task training.
Across 47 benchmark evaluations spanning 8 domains, Ling-3.1-flash-base outperformed DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base on 32 benchmarks, achieving top scores. Key evaluations across knowledge understanding, logical reasoning, coding, mathematics, and long context correspond to the foundational capabilities personal assistants require for managing complex tasks, serving as a critical launchpad for sophisticated workflows. Ling-3.1-flash-base is available to developers with BF16 base weights under the MIT license, enabling continued pre-training, domain adaptation, and post-training tailored to proprietary knowledge, workflows, and tools.

Building on top of these base capabilities, Ling-3.1-flash and the subsequent 1019 release use post-training to continuously reinforce reasoning, tool usage, and multi-step execution, progressively boosting performance in real-world workflows. From the Base model to post-trained checkpoints, Bailing’s R&D focus has steadily expanded toward goal adherence, tool utilization, and autonomous task progression.
Furthermore, to address multimodal processing needs in personal workflows, Bailing trained the vision-language model Ling-3.1-flash-VL on top of Ling-3.1-flash. It layers visual comprehension on top of established reasoning, planning, and tool-use capabilities; it is currently deployed in Lingguang and will be publicly released in the future.
Derived from Real Tasks, Deployed to Real Applications
Bailing is systematically refining the model capabilities demanded by Personal Agents. These model-level advancements are being validated and explored in live applications. Ling-3.1-flash has undergone alignment training based on user requirements and feedback from applications like Lingguang and Afu Doctor, continuously advancing its professional expression and communication skills.
“Derived from real tasks, deployed to real applications.” Bailing stated that through a continuous feedback loop between model R&D and practical deployment, it aims to enhance the model’s ability to comprehend goals, maintain task state, and adapt to changes, while collaborating with the developer community through open source to explore further application possibilities.
Rather than merely chasing parameter scale, Bailing places greater emphasis on integrating foundation model capabilities into real workflows to accomplish tangible tasks. Since the second half of this year, Bailing has accelerated its open-source releases and real-world scenario iterations.
Previously, based on the next-generation foundation model Ling-3.0-flash, Bailing introduced the finance-enhanced Ling-3.0-flash-Fin and the healthcare-enhanced Ling-3.0-flash-Sante, translating foundational capabilities into specialized professional domains. Among them, Ling-3.0-flash-Fin (124B total parameters, 5.1B activated) integrates AI directly into investment research workflows—connecting the entire pipeline from information gathering and data analysis to research report generation—rivaling 3T-scale frontier models on high-difficulty investment research tasks and delivering industry-leading performance among open-source financial agents. Meanwhile, Ling-3.0-flash-Sante brought high intelligence-to-compute efficiency to healthcare, achieving top results among sub-1.6T models with just 124B parameters and delivering overall performance on par with multiple flagship models.
This article was provided by Ant Group and reprinted with permission by QbitAI. The views expressed belong to the original author.
All rights reserved. Reproduction and use in any form without authorization is strictly prohibited.
- Bridging Energy and Compute: Damao Tech Defines Power-Compute Synergy with Energy Brain | CPE 20262026-10-27
- OXR Goes Open-Source! 1.3B Model Sweeps #1 on Three Global Leaderboards, Outpaces Paddle and MinerU with RSI Raising the Bar2026-10-27
- 2026 Global Developer Pioneers Conference and International Embodied AI Skills Competition Successfully Concludes2026-10-27
- 30B LLMs Land on Smartphones as Inference Engines Begin Self-Evolution2026-10-24
Scan QR code to share to WeChat Moments
Related Reading
Say Goodbye to Complex Prompts! Ant’s New Approach Enables AI to Automatically Understand Personalized Needs
DeepSeek personalized alignment improved by 15%
Wen Le2025-08-03
Ant’s Medical LLM Takes Double Championship on MedBench Benchmark with Native Multimodality and Hundreds of Billions of Tokens
High scores of 97.5 on official benchmark and 98.2 on self-test benchmark
Bai Jiao2025-03-12
Another Domestic Decision-Making Model Released! Accuracy Surpasses Models from Microsoft, Jev, and Others
On October 22, Ant Digital Technologies officially released the decision model LingDT-Decision-1.0.
QbitAI2026-10-22
Alipay Medical LLM Unveiled! Beats GPT-4 on English and Chinese Medical Exams, Already Deployed in Top Hospitals across Jiangsu, Zhejiang, and Shanghai
Recognition accuracy for medical reports, pharmaceuticals, hair, and other images exceeds 90%.
Bai Jiao2024-07-08
Understanding Human Nuance and Flawless Execution: Ant’s Trillion-Parameter Open-Source Model Maxes Out EQ and Agent Capabilities
A trillion-parameter heavyweight that moves with agile grace
Kresi2026-02-19
Ant Digital Technologies Announces Open-Sourcing of Data Analysis Agent Technology, Currently Topping BIRD
QbitAI2025-12-15
Trending Articles
Tracking new AI trends, reporting on tech industry breakthroughs
QbitAI Copyright © Beijing Geek Partner Technology Co., Ltd. Beijing ICP Filing No. 17005886-1
