The jev-visual tool has been released for local vision on Apple Silicon

Developers introduced jev-visual — a project for performing computer vision tasks on devices with Apple Silicon chips. The tool uses the Qwen3.5-0.8B architecture in conjunction with the MLX framework to provide high-speed local inference.

The system works based on the direct scoring method, which allows the model to interpret visual data without the need to access cloud APIs. In the current version, support for real-time video stream processing is implemented, which is demonstrated by examples of automated object sorting and a simplified version of the Breakout game.

Using 0.8B parameters allows the model to run on devices with limited RAM while maintaining low latency. Integration with MLX provides optimization of calculations for the Apple architecture specifics, minimizing the load on the central and graphics processors when performing recognition tasks.

The project is aimed at developers creating local AI agents that can interact with the visual environment. The source code is available on GitHub under an open license, which allows integrating the module into third-party applications for automating processes that require real-time image analysis.