Ant Group's Bailing Open-Sources Ling-3.1-flash Models for Personal Agents

Lingguang 2.0 and the Personal Agent Evolution

On October 27, Ant Group's Bailing team released Lingguang 2.0, upgrading it to a "Personal AI Assistant that can handle tasks." The assistant proactively follows up on tasks and continuously manages complex personal affairs such as trip planning (including weather checks and appointment reminders) and Double 11 shopping price comparison with continuous monitoring and alerts. Within user-authorized scope, Lingguang understands personal affairs, arranges follow-ups, and proactively reminds at key nodes. It relies on cross-session memory, long-horizon task scheduling, and an independent cloud host so tasks continue running even after the user exits the conversation.

Open-Source Model Releases

Accompanying the launch, Bailing open-sourced two models:

Ling-3.1-flash-1019 – a post-trained model scoring 44 on the Artificial Analysis Intelligence Index v4.3.2.

Ling-3.1-flash-base – the base model released under MIT license with BF16 weights, enabling developers to continue pre-training, domain adaptation, and post-training.

Both models are available on Hugging Face and ModelScope:

Ling-3.1-flash-1019: Hugging Face: https://huggingface.co/inclusionAI/Ling-3.1-flash-1019, ModelScope: https://modelscope.cn/models/inclusionAI/Ling-3.1-flash-1019

Ling-3.1-flash-base: Hugging Face: https://huggingface.co/inclusionAI/Ling-3.1-flash-base, ModelScope: https://modelscope.cn/models/inclusionAI/Ling-3.1-flash-base

Base Model Architecture and Benchmarks

Ling-3.1-flash-base uses a hybrid linear Mixture-of-Experts (MoE) architecture with approximately 560B total parameters and 25B activated parameters per token . The design balances capacity, long-sequence handling, and single-step compute efficiency to support further training for real-world tasks.

In base model evaluation across 8 domains and 47 benchmarks , Ling-3.1-flash-base achieved the highest score among three compared models on 32 benchmarks . Key areas include knowledge understanding, logical reasoning, code, mathematics, and long-context — capabilities directly relevant to personal assistants handling complex affairs.

Base model benchmark results (source: model card). The "32 highest" claim refers only to the three models in the table.

Post-Training: From Ling-3.1-flash to Ling-3.1-flash-1019

After establishing base capabilities, post-training focused on teaching the model to organize and apply those capabilities around user goals. Ling-3.1-flash was strengthened for general-purpose agents, search, office productivity, and software development, emphasizing reasoning, tool use, and multi-step execution. Evaluations used GDPVal-AA V2.1 (work tasks), FrontierSWE (software engineering), and Terminal-Bench 4.0 (terminal operations) to test real-workflow performance.

Ling-3.1-flash-1019 (released October 22) adds further post-training optimization, improving reasoning and agent abilities in real scenarios. Compared to the previous Ling-3.1-flash:

Comprehensive intelligence index: +3 points to 44

AA-Briefcase Elo: ~1400 → 1524

AutomationBench-AA: 62% → 68%

Scientific reasoning and scientific coding benchmarks also improved.

Performance comparison: Ling-3.1-flash vs Ling-3.1-flash-1019 (data from model card; Elo may update with leaderboard).

The gains in work-task and automation benchmarks provide quantitative evidence for further iteration. For a Personal Agent, these capabilities must translate into reliable judgments after each information update and concrete results during continuous execution.

Vision-Language Model: Ling-3.1-flash-VL

Personal affairs often involve photos, screenshots, and software UIs. Building on Ling-3.1-flash, Bailing trained Ling-3.1-flash-VL , adding visual understanding while retaining reasoning, planning, and tool-use skills. The model is already deployed in the Lingguang product and will be released publicly later.

A concrete example from the Lingguang service: a parent uploads a homework photo; the assistant verifies questions and solutions, extracts knowledge points. The user then asks to "consolidate in two weeks," turning a one-off homework check into a phased practice schedule with daily reminders and learning records. The system generated a two-week plan, task pages, and record pages, linking subsequent exercises to the same goal.

This scenario requires both vision and continuous execution: first understand the photo content, then organize follow-up actions around the learning goal; as new homework or feedback arrives, continue advancing within the existing plan. Vision lets the assistant ingest more real-world information, while existing reasoning and tool capabilities convert that information into next steps.

Development Philosophy: From Real Tasks to Real Applications

Personal Agent improvement relies on a loop between model training and real application. Ling-3.1-flash has undergone alignment training using feedback from Lingguang and the "Aifu Doctor" scenario, improving professional expression and communication. From base model design (Ling-3.1-flash-base) through post-training iterations (Ling-3.1-flash, Ling-3.1-flash-1019) to the upcoming vision model (Ling-3.1-flash-VL), Bailing is progressively completing the capability stack for Personal Agents. By open-sourcing models, they invite developers to explore more possibilities together.

From real tasks, to real applications. Bailing aims to enhance the model's ability to understand goals, maintain state, and adapt to changes through continuous cycles of R&D and practical deployment, making AI a reliable, handy assistant for daily affairs.

Additional Leaderboard Information

Further benchmark charts are provided in the original article, showing detailed comparisons across multiple evaluation suites.