Launch of Kev Decision Models Enhances Cybersecurity Capabilities

• •Kev-9B achieves the highest accuracy among the new models with 0.812 out-of-domain.
• •Kev-4B and Kev-0.8B also show significant improvements over previous versions.
• •The models utilize advanced LoRA adapters and pointer heads for enhanced decision-making.

On September 21, 2026, three new decision models named Kev-4B, Kev-9B, and Kev-0.8B were released, each designed to process documents and answer typed questions with high accuracy. Kev-9B is noted for its superior performance, achieving an out-of-domain accuracy of 0.812 on the development partition, while Kev-4B and Kev-0.8B also showed significant improvements over their predecessors. These models utilize LoRA adapters and pointer heads on the Qwen architecture, with varying numbers of trainable parameters: 33.8M for Kev-4B, 45.4M for Kev-9B, and 11.3M for Kev-0.8B. The models are intended for use in TypeSafe's public /v1/systemone contract, enhancing decision-making capabilities in cybersecurity applications. The release indicates a focus on improving accuracy and calibration in decision-making models, with Kev-9B achieving the lowest Brier score of 0.243. Each model's performance has been validated against specific datasets, demonstrating their effectiveness in various scenarios.