Ant Group open-sources Ling-3.1 Flash for persistent agents
Ant Group’s Bailing team released Ling-3.1 Flash 1019 with stronger long-horizon reasoning, tool use, memory, and workflow automation.
Ant Group’s Bailing foundation-model team released Ling-3.1-flash-1019 and the companion Ling-3.1-flash-base on October 28, targeting personal agents that can continue working across sessions rather than simply answer prompts.
What changed
The 1019 version is a post-trained upgrade focused on multi-step execution, tool use, and sustained task progress. Bailing reports a score of 44 on Artificial Analysis’ Intelligence Index v4.3.2, three points above the previous version. It also reports gains on knowledge-work, automation, scientific reasoning, and coding-related evaluations, including AutomationBench-AA, where the score rose from 62% to 68%.
The base model is released under the MIT license for developers that want to continue pre-training, adapt the model to private workflows, or perform their own post-training. Bailing says Ling-3.1 Flash has 560 billion total parameters, activates roughly 25 billion per token, and supports a context window advertised at up to one million tokens. A vision-language variant is being used internally in Ant’s Lingguang personal-assistant product but has not yet been released.
Why it matters
The release is important because it ties an open-weight model to a concrete Asian product cycle: Lingguang’s recent upgrade added cross-session memory, background task scheduling, and an independent cloud environment. Those features require a model to maintain goals, recover from intermediate failures, and decide when to use tools. The public weights therefore give developers a candidate foundation for persistent personal and business agents, not merely another chat model.
The main caveat is reproducibility. The strongest benchmark numbers come from the model maker, while the 1019 checkpoint is large enough to make local deployment demanding. Its practical value will depend on inference support, long-running reliability, and whether independent users can reproduce the claimed gains outside Bailing’s own application stack.