Qwen-AgentWorld-35B-A3B: a local 'world model' you can run at home Open Models (vettedconsumer.com)

On June 22, Qwen introduced the Qwen-AgentWorld-35B-A3B, a new type of local language model that functions as a 'world model' rather than a typical chatbot or coding assistant. Unlike standard models that predict the next action, the 35B-A3B predicts the next state of an environment based on given actions and current conditions. This capability allows it to support a variety of tasks across seven different agent domains, such as software engineering, terminal commands, and web interactions. The model is designed to be run on consumer-grade hardware, requiring around 21 GB of GPU memory, but operates efficiently due to its innovative architecture featuring Gated DeltaNet and gated attention blocks, which optimize memory usage during long context runs. This launch is significant for the AI/ML community as it showcases a practical application of world modeling, allowing developers to create more complex and contextually aware agent systems without relying on expensive API calls. The model boasts a large context window of 262,144 tokens and a mixture-of-experts architecture that enables effective long-agent trace processing. While early reactions are mixed, with some users heralding its potential for streamlined workflows and others questioning its classification as a 'world model,' it offers a promising tool for building responsive and efficient AI agents. As developers start utilizing this model, clearer insights into its effectiveness and integration within existing frameworks should emerge.