The Pulse

Nace Releases Drex 1.5’s 9B Weights Under a Restricted License

Nace.AI published Drex 1.5, a 9B decision model that returns probabilities for supplied choices rather than generating text. Its weights and inference code are available on Hugging Face, but a modified Open RAIL-M license limits some commer

AI.info Team · Written with AI · 2026-10-10

A decision model from Nace.AI: it reads a state (text or JSON) and typed questions (choice, noul, ordinal score) and returns a probability for every option in one forward pass per question.

Nace.AI’s Drex 1.5 model card

Nace.AI published Drex 1.5’s weights on Friday, October 9, alongside code for running the model locally. The 8.95-billion-parameter model is built to score decisions rather than compose answers: developers provide a text or JSON state and typed questions, and Drex returns probabilities across the options they supply. The model card says it generates no text.

Drex scores choices instead of writing replies

The model supports three question types: choice for selecting among labels, noul for a yes-or-no probability, and ordinal score for an ordered scale. Its separate pointer head scores options using the backbone’s hidden states. Since the possible answers come from the caller, Drex cannot add a new option to the list.

That design targets components such as routing, classification and tool selection, where software needs a bounded answer rather than a paragraph to parse. Nace packages an inference script and a local HTTP server that accepts requests through a /v1/systemone endpoint. The model card also documents running a quantized version through llama.cpp or Nace’s Ollama fork.

The 58.08 score comes with a close comparison

Nace reports a score of 58.08 on the public portion of Decision Index 0.3.1, which covers 37 benchmarks across five areas. The model card places Jev 1.13.0 at 57.96 and Bespoke Nimble 9B v3 at 57.19; it says Drex, Jev and Nimble fall within the leaderboard’s 0.9-point tie band. Nace says Drex leads Jev on 20 of the 37 benchmarks.

The card identifies the Drex result as Nace’s own run of the benchmark kit, while the comparison rows for other models come from the leaderboard. On a separate 231-item JevBench evaluation, Drex scores 86.2%, below Jev’s 87.0%; both score 73.9% on the hard tier. Nace also reports 93.4% accuracy on long-document questions using 32,000 to 128,000 tokens, with a median latency of 2 seconds.

Local use requires substantial memory—and a license check

The bf16 weights take about 18 GB, and Nace says its Python inference path requires a CUDA GPU. A Q8_0 GGUF version is about 9.5 GB and is documented for Apple silicon and CPU use through supported tools. The model’s context limit reaches 131,072 tokens, though the model card lists 16,384 as the default.

“Open” comes with defined limits here. The weights use Nace’s modified Open RAIL-M license, not an unrestricted license: it restricts certain uses and says commercial use by organizations above specified revenue or funding thresholds requires a separate written license. For developers evaluating Drex, that makes the model’s published weights and local-running instructions only part of the deployment decision; the license terms also shape who can use it and how.