Nace.AI has released Drex 1.5, a 9B parameter decision model designed to score specific options rather than generate text. The model utilizes the MiMo-V2.6-Distill-Qwen-9B architecture, featuring 32 layers with hybrid attention and a dedicated pointer head for scoring. It operates via a single forward pass, supporting choice, yes/no, and ordinal score question types.
The model is compatible with the POST /v1/systemone API, allowing existing clients of the Jev model to integrate it by updating environment variables. Deployment is supported through custom forks of Ollama and llama.cpp. Nace.AI confirmed consistent performance across AWS g5.2xlarge instances and Apple M5 Pro hardware.
In performance testing on Decision Index 0.3.1, Drex 1.5 performed within a 0.9-point tie band alongside Jev and Nimble. While it outperformed Jev on 20 of 37 benchmarks, it showed lower results in knowledge-heavy tasks like GPQA Diamond and MMLU-Pro. The model is available under a RAIL-M license with specific usage restrictions.
Drex 1.5 can be integrated into various development tools, including Claude Code, Cursor, and GitHub Copilot, via a dedicated agent skill. Nace.AI is currently offering a $25 sign-up bonus for new cloud users. The model is strictly a decision layer and does not possess general text or code generation capabilities.
