Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon

Nace AI has published open weights for Drex Max v1, the decision model it previously offered only as a hosted service. It is a dense 27-billion-parameter model that reads text, JSON or images, takes typed questions, and returns a probability for every option in a single pass. It does not write prose.

The practical news is where it runs. Nace ships full-precision weights for NVIDIA GPUs, an 8-bit GGUF for Macs with 48 GB, a 4-bit build aimed at Apple Silicon machines with 24 GB or more, and forks of llama.cpp and Ollama. The local Mac and CPU paths are text only for now. The license is Nace's Open RAIL-M, the same family as its earlier open Drex releases.

For teams that route agent actions or screen content against policy, this means those calls can stay on their own hardware instead of going out to a vendor on every request.

We like this release more than the hosted one. A model that answers only from a fixed menu, with a probability attached, is easy to reason about. Putting the weights on a laptop makes it easier still: a security team can test it offline, log every decision, and keep sensitive tickets or customer records off someone else's servers. For guardrails especially, local and inspectable beats remote and opaque.

Nace also did something we want to see more often. It reports that its quantized builds stay close to the full-precision model, with the 8-bit file on an M5 Pro landing within 0.02 per probability of an H100 run and the 4-bit build producing no flipped choices in its tests. That is the right question to answer, because a guardrail that quietly changes its mind when compressed is worse than none.

Now the caveats. Every benchmark is self-reported, including 67.52 on Decision Index 0.4 against 61.25 for Drex v2.0. Nace says version 0.4 uses the same public suite as 0.3.1, but cross-version comparisons deserve an independent check. The model is strongest on tool use and weakest on knowledge, reasoning and arts, so it is a fast judge for narrow calls, not a thinker. And the coverage announcing it was sponsored by Nace, which is reason enough to wait for outside numbers.

The harm to name is quieter. When policy checks become cheap and local, organizations will run them everywhere: on support tickets, on job applications, on flagged accounts. A confident probability looks objective. If nobody owns the threshold or reviews the borderline cases, automated triage hardens into automated denial, with no one able to say why.

Our read: a genuinely useful open release that moves decision models closer to the people who have to answer for them. I'm watching whether builders treat the probabilities as evidence for a human, or as the final word.

Decision models return probabilities over developer-defined answers rather than generating text. Nace has open-sourced smaller Drex models before, including a 9B Drex 1.5 and a mixture-of-experts Drex 2.0. Drex Max launched as a hosted-only model; this release opens its weights.