Nace.AI drops Drex sub-6B decision model

Using heavyweight generative LLMs for discrete operational choices is wasteful, making lightweight, single-pass decision models like Drex a far more sensible architecture for agentic control loops.

• –Direct probability output: By bypassing generative autoregression and text parsing, Drex evaluates choices in one forward pass, reducing latency and slashing token consumption by 5x.
• –Efficiency over scale: Ranking first on Decision Index 0.2 with under 6B parameters shows that task-specific modeling can outperform models four to six times its size.
• –Practical edge and local deployment: Open weights enable developers to run fast, deterministic routing directly inside local pipelines, game logic, or robotics without vendor lock-in.
• –Narrow benchmark margin: A 0.06 score lead over Jev is thin, meaning developer adoption will hinge on whether single-pass probability outputs generalize well across complex custom schemas.