Drex

Drex is a family of decision models from Nace.AI, a Palo Alto company founded in 2024. Instead of writing text, it takes the facts of a case (text, screens, images or video) along with typed questions and returns a probability for every option in a single pass. It is available through a hosted API and console, and Drex 2.0 ships as open weights.

What It Is

Drex is a decision model rather than a text-generating LLM. You send a case and typed questions: a choice between named options, a yes/no question, or a score on a scale you define. Each comes back with probabilities, so Drex can only answer with the options you supplied. Several questions about one case can share a single pass.

The family has three models. Drex Max v1.1 (drex-max) decides from screens, images, charts and video and is aimed at agents that use computers and browsers. Drex 2.0 (drex-v2.0) is an open-weights model that Nace says fits on a single accelerator. Drex 1.5 (drex-v1.5) is a lean text model for high-volume routing, triage and classification.

Where It Fits

Nace positions Drex for browser-using agents that pick the next click from a screenshot, for routing and classification, and for finance, banking and legal decisions such as sentiment in finance news, banking intent routing and contract analysis. Drex for Search and Drex for Finance are listed as use-case pages. The API is described as wire-compatible with Jev, Nace's earlier decision model, so switching takes three environment variables.

Benchmarks and Performance Claims

Nace reports that Drex Max v1.1 scores 70.8 on the public Decision Index (edition 0.4.0), against 60.7 for Jev 1.14. It also reports a median API latency of 172 ms and a cost of $0.03 per thousand decisions. These are vendor-published figures. Nace states that Drex is trained on the official training splits of benchmarks and evaluated only on held-out splits.

Deployment Model

Drex can run on Nace-managed infrastructure, in a customer's own cloud account, on-premises or at the edge. Drex 2.0 has 35B parameters with about 3B active per token, a 128K-token context, and bf16 weights of about 70 GB (about 37 GB for the 8-bit build). Nace also offers tuning on a customer's labelled outcomes. The company states it undergoes SOC 2 Type II audits.