Nace AI open-sources Drex 2.0 decision model weights

Drex 2.0 is a small model from Nace AI built for the routine choices inside an agent's workflow, such as whether to approve an invoice and where to send it, rather than for writing code or prose. Given the relevant state and a set of questions, it returns what Nace calls a calibrated probability for every option in one pass. In the demo on Nace's developer page, it puts 93% on approving one invoice and 88% on routing it to payments. Nace has now released the weights, linked from GitHub and Hugging Face, so developers can run the model themselves instead of only through Nace's paid API. The announcement's "94% sure" is the model's answer to whether Nace should open-source it, not an accuracy figure.

Nace's case for the model is speed and cost. It argues that paying a frontier model for each of the thousands of daily steps in a workflow is slow and expensive, and its chart puts Drex 2.0's median API time at 151 milliseconds per decision, against about 2.0 seconds for GPT-6 Sol and 3.5 seconds for Claude Opus 5.5, at $0.06 per 1,000 decisions against $3.33 and $10.17. Those two frontier models have no score on the chart's Decision Index, a chance-corrected measure of these choices, so the page does not show whether Drex decides as well as they do. Against Jev 1.14, a decision model some of Nace's quoted users switched from, Drex 2.0 scores 61.3 to 60.7 and responds faster, but costs 1.4 times as much through the API. All of these figures come from Nace's page.