Nace AI opens up Drex, a 183ms decision-making model family

Most automated business workflows don’t need a chatbot to write poetry. They just need a fast, reliable pick from a fixed set of choices.

Nace AI just published a 51-page technical report breaking down Drex, a family of open-weight decision models purpose-built for that exact job.

Built for one job: picking the right option

Drex v1.5 was trained on 1.15 million verified decision examples, tackling tasks ranging from contract analysis to flagging phishing emails.

A second stage uses a larger model to train Drex via RLAF, calibrating its confidence scores so that its predicted percentages actually reflect real-world accuracy.

In testing, Drex Max flagged a suspected duplicate invoice for less than a penny.

Nace reports a median response time of just 183 milliseconds, compared to roughly 2 seconds for GPT-6 Sol.

Weights and code for Drex Max, v2.0, and v1.5 are now up on Hugging Face for local hosting, with a free web test interface available on Nace Console. However, Drex Max 1.1 and Drex for Finance remain closed-source for enterprise clients.

Why it matters

We think Nace is smart to target this narrow slice of enterprise AI. For high-throughput pipelines like fraud detection or compliance triage, full-sized foundation models are often overkill and far too sluggish. Dropping latency to 183ms while outputting calibrated probabilities makes local automation genuinely practical.

The tradeoff is obvious: Drex can only pick from predetermined answers, so it won't help with generative or open-ended reasoning.

Still, if your system just needs to make quick, auditable choices on a tight budget, this is well worth spinning up locally.