Developers introduced the AgentJev model for agent cycles

Developers have released AgentJev-0.6B — a specialized language model designed to work as part of agent systems. The release is focused on performing tasks that require high-speed decision-making without generating long text sequences. The architecture is based on the Qwen3-0.6B model, fine-tuned to work in zero-token-decoding mode. Instead of standard text output, the system returns probability distributions for typed decisions, such as logical values (boolean), choices, and numerical scores. This approach allows the model to be integrated as a "reflex" node in agent cycles. The tool demonstrates performance at the level of 76 points when performing tasks for 21 hours of testing. The model is optimized for scenarios where an agent's instant response to input data is required without the need for token decoding, which reduces delays in reasoning chains. The project repository is available on GitHub under an open license. Using AgentJev allows developers of agent systems to replace heavy LLMs with a lightweight module for classification and making discrete decisions, which significantly accelerates the work of autonomous performers in real-time.