Ant Group releases Ling-3.1-flash: agent runs for 10+ hours straight

Ant Group has launched Ling-3.1-flash, a new fast model in the Bailing (百灵) family. Chinese outlet Jiqizhixin reports that it can work on a single task for more than ten hours without a break, and that in demonstrations it was asked to build cities. The release's main message: the model is designed for real workflows, not for impressive benchmarks.

Context: first Jev, then Ling

A month earlier, an agent called Jev became popular in the Chinese AI community. It can barely write long texts. Its job is something else: making quick decisions inside an "observe, act" loop. That sounds modest, but it proved useful in practice, and many developers have plugged it into their agentic pipelines. Its most visible demonstration: Jev defeated the Ender Dragon in Minecraft entirely on its own, starting from scratch.

The outlet presents Jev's story as a symptom of this year's shift. Attention is moving away from models that answer nicely in a chat and toward those that do work reliably and cheaply as part of a chain with other components. Ling-3.1-flash fits that logic: in Chinese labs' model lines, the "flash" suffix usually signals a priority on speed and inference cost.

What we know about the release

The source material available to us cuts off after the introduction, so architecture details, model size, pricing and benchmark results are not yet confirmed. Below is only what follows from the headline and the opening section:

• developer: Ant Group, Bailing (百灵) family;
• model: Ling-3.1-flash, a lightweight, fast version of the line;
• claimed autonomous operation: more than ten hours straight;
• demo scenario: generating and building cities;
• positioning: integration into real workflows and agent chains.

The "more than ten hours" figure deserves separate attention. Most agents break down much sooner: the context bloats, errors accumulate, and the model loses its goal or gets stuck in a loop. If Ling-3.1-flash really holds its course over that distance, it is a strong argument for tasks such as refactoring large codebases, long testing cycles and processing large document sets.

In practice: what to check before adopting it

Until independent measurements appear, the ten-hour claim is best treated as marketing, but verifiable marketing. Developers considering the model for agentic systems should:

• run their own long scenario and measure the share of successful runs, not just average run time;
• look at the cost per hour of agent operation: for flash models, this is usually what decides whether autonomy pays off;
• test behavior when tools fail: whether the model can roll back and retry steps rather than continue with a corrupted state;
• compare it with narrowly specialized agents like Jev, which win through decision speed.

Once Ant publishes a technical report and access terms, the picture will get clearer. For now, the release shows where competition is shifting: the winner is the model that can last a long shift without supervision.