The compact NanoJev model is presented for managing game agents

Developers presented NanoJev — a specialized language model with 0.6 billion parameters, designed for decision-making in game environments. The architecture of the novelty is based on Qwen3 and optimized for performing tasks in Maze, Snake and ViZDoom. During testing, the model demonstrated superiority over the original version of Jev and the basic Qwen model without additional tuning. The main emphasis during training was placed on improving the accuracy of decision-making in conditions of limited computational resources, which allows the system to be launched on devices with a low amount of RAM. The technical characteristics of NanoJev include the use of specialized weights adapted to the specifics of game mechanics. The model shows stable results in navigation and character control tasks, where high-speed processing of input data in real-time is required. The source code and model weights are already available for download in the repository on the Hugging Face platform. The release of NanoJev expands the possibilities for creating lightweight AI agents that can function in game simulations without resorting to cloud powers. The use of the Qwen3 architecture as a foundation provides a balance between compactness and context processing quality, which is critical for autonomous control systems in dynamic environments.