Developers introduced the compact NanoJev model for decision making

Researchers have released NanoJev — a specialized model with 0.6B parameters designed for real-time decision-making tasks. The system's architecture supports parallel data processing, allowing it to generate full probabilistic distributions without the need for classical autoregressive decoding.

In testing, the model demonstrated high efficiency in game environments, including labyrinth simulations and the game "Snake". According to benchmark results, NanoJev outperforms the Jev and Qwen models in similar scenarios. The system's speed was rated at 9 points on an internal evaluation scale with training or inference time of 10 hours.

The technical implementation of the model is focused on minimizing computational costs while maintaining prediction accuracy. The abandonment of step-by-step token decoding allows the system to respond faster to environmental changes, which is critical for agent control tasks in dynamic conditions.

The release is available in open access on GitHub under a free license. The developers note that the small size of the model allows it to be run on consumer equipment without using powerful server GPUs, which simplifies integration into local AI systems and autonomous game engines.