Drex v1.5

Not yet independently verified. The release date of 28 September 2026 rests only on the creation of the Hugging Face repository; the vendor's own announcement date for v1.5 was not found, and its blog post of 24 September 2026 introduced the first Drex. The benchmark figures are Nace AI's own, so no score is stored. The licence is a modified Open RAIL-M, not an open source licence approved by the OSI. No independent report was found. We will update this when it can be confirmed, and remove this note.

Drex v1.5 is an open-weight decision model from Nace.AI. Given a state and a set of typed questions, it returns a probability for every option supplied rather than generating text. Its Hugging Face repository was created on 28 September 2026, and the weights are also listed in Nace's public GitHub repository for Drex decision models.

What Drex v1.5 does

The model takes text or JSON and answers three kinds of question: a choice between named options, a yes or no question that returns the probability of yes, and a score on an ordinal scale. Each question needs one forward pass, and the model cannot return an option it was not offered. It has about 9 billion parameters and is built on a distilled Qwen 3.5 9B backbone, with a pointer head that scores the options. Context defaults to 16,384 tokens and can be raised to 131,072.

It can run through a Python server on NVIDIA GPUs, a llama.cpp fork, or a Nace fork of Ollama. A hosted Drex API serves the same request format.

Nace reports a score of 58.08 on the public Decision Index 0.3.1, which it says ranks first among models under 10B parameters, with the leading three within a 0.9 point tie band. These benchmark figures are the vendor's own, mostly from its own runs.

Who it is for

It suits developers who want classification, routing or approval decisions from a small model they can host themselves. The page notes that the local server has no authentication and binds to the local machine by default.

Key facts

• Maker: Nace.AI
• Repository created: 28 September 2026
• Licence: Nace.AI Open RAIL-M for the weights; the runtime code is Apache-2.0
• Pricing: open weights; the hosted API has a free key
• Platforms: CUDA GPUs for the Python path; llama.cpp and Ollama for CPU and Apple silicon
• Weights: Hugging Face, nace-ai/drex-v1.5