Yandex, Together AI: Qwen 3.x runs async agents, no training

The framework is called AsyncLLM, and its central claim is that any recent open-weights LLM can be used as an asynchronous agent without extra training. That is the pitch in "LLMs are General Asynchronous Agents," a preprint from seven researchers affiliated with Yandex, Together AI, HSE University and the Yandex School of Data Analysis.

The authors' starting complaint is that today's agents run on a strict Thought-Action-Observation loop, and that "many real-world use cases are not sequential: voice assistants, embodied agents, and monitoring systems receive new inputs while they think or perform another task."

Their answer is an "asynchronous LLM framework that lets users (or the agents themselves) define inference coroutines with overlapping memory states," modeled on Python's async/await pattern. The paper reports that Qwen 3.x models operate this way for streaming video understanding, videogames and monitoring "without task-specific training," resting on the hypothesis that because LLMs learn human reasoning patterns during pretraining "they may be able to imitate human asynchrony with proper framing."

Code is published at github.com/dvmazur/async_llm. The abstract does not report head-to-head numbers against a sequential baseline, and only the Qwen 3.x family is named among tested models; voice assistants, one of the motivating examples in the introduction, are not among the three demonstrated task families.