On 30 October 2026, Nace.AI released the weights of Drex Max, a 27-billion-parameter "decision model" that runs on Apple silicon Macs with 24 GB of memory or more. A decision model does not write prose. You give it some text and a question with fixed options, and it returns a probability for each option. That suits jobs such as sorting support tickets, flagging contract clauses or screening forms. Two catches matter before you download it: the Mac version reads text only, and the license requires a separate written commercial license once your organization passes $1 million in revenue or funding.
What a decision model does
An ordinary chat model writes an answer word by word, and you then have to read or parse it. A decision model answers in a single pass with numbers:
| You send | Drex Max returns |
|---|---|
| A ticket: "I was charged twice for the same order." Question: which team should handle it? Options: billing, technical, other | "choice": "billing", with billing at 0.9988 probability and a confidence of 0.9982 |
That example is the one on Nace's model card. Because the output is a probability, a script can route confident answers automatically and send uncertain ones to a person.
Drex Max supports three question types, the same ones Ollama's built-in decision models use: choice (pick one option), noul (a yes/no probability) and score (a rating on an ordered scale). Our guide to Ollama decision models walks through a complete request and how to read each field.
The model in brief
| Item | Drex Max v1 |
|---|---|
| Size | 27 billion parameters (dense: every parameter is used for every answer) |
| Output | A probability for every option, from one pass through the model |
| Context length | 16,384 tokens per request by default; up to 262,144 with the DREX_MAX_CONTEXT setting |
| Input types | Text and images through Nace's Python code on an NVIDIA GPU; text only on Macs and through llama.cpp |
| License | Nace.AI Open RAIL-M, with revenue and funding limits (see below) |
| Hosted versions | Also offered through Nace's API, DeepInfra and OpenRouter |
Nace reports a score of 67.52 on its own public Decision Index 0.4, and says Drex Max ranks fourth of 18 models on the Artificial Analysis Decision Index with a score of 64. These are Nace's figures; Local LLM Guide has not tested them.
How much memory it needs
Nace's model card lists these routes:
| Version | Memory | Where it runs |
|---|---|---|
| Full precision (bf16) | About 49 GB | NVIDIA GPU, through Nace's Python code |
| 8-bit core model | About 26 GB | NVIDIA GPU, through Nace's Python code with --8bit |
| 8-bit GGUF (Q8_0) | 48 GB listed; the file itself is about 27 GB | Apple silicon, Linux or Windows, through llama.cpp, Kapitsa or Nace's Ollama build |
| 4-bit (affine Q4) | Macs with 24 GB or more | Apple silicon, through MLX or Kapitsa |
So an Apple silicon Mac with 24 GB of unified memory is the entry point, using the 4-bit version. A 16 GB Mac is below Nace's listed minimum. If you are deciding between Macs, see the best local AI setup for Mac.
On a Mac, Drex Max reads text only. The card states that images are not supported in the GGUF, Kapitsa or MLX versions; image questions need the Python route on an NVIDIA GPU.
Running it on a Mac
Drex Max is not in the regular Ollama library. You cannot run ollama pull drex-max. The card describes four local routes, each set up from the command line:
- MLX, Apple's machine-learning framework: download the 4-bit version (
nace-ai/drex-max-v1-affine-q4) and run Nace's MLX script in a Python environment. This is the most direct Mac route for the 24 GB tier. - Kapitsa, Nace's own Mac runner, installed with Homebrew (
brew install nace-ai/tap/kapitsa). It serves the 8-bit GGUF file locally. - llama.cpp, using a Nace branch (
drex-max-v1) that you compile yourself. - Ollama, using Nace's modified copy of Ollama, which you also compile yourself. It answers on the same
/v1/systemonedecision endpoint as the official Ollama.
For a text-only Mac test using MLX, download Nace's affine Q4 package and run its supplied request. For llama.cpp, use Nace's drex-max-v1 branch. Its ready-made Q8_0 file is 27.25 GB and avoids downloading the BF16 source just to convert it. If you are new to these tools, the GGUF vs MLX comparison explains the two model formats.
An easier start: decision models already in Ollama
If you want to try local decision models without compiling anything, the official Ollama app already includes several. Ollama 0.40.7 (18 October) added pplx-decider, a 27B decision model from Perplexity, and Ollama 0.40.11 (24 October) added the d3 decision models in 2B, 4B, 9B and 27B sizes. These install with an ordinary ollama pull and use the same request format. Check each model's Ollama page for its download size, and make sure the tag you choose runs on your computer rather than in Ollama's cloud. Start there, and move to Drex Max if its published scores matter for your task and your Mac has the memory.
The license: read this before using it at work
Drex Max uses the Nace.AI Open RAIL-M license. It is open in the sense that the weights are free to download, but it carries conditions that matter to businesses.
Who needs a commercial license. The free terms do not cover commercial use by a person or organization that, in the prior 12 months, generated more than $1 million in gross revenue, or that has raised more than $1 million in total equity or debt funding. The test counts the user's employer and affiliated entities, so an employee of a larger firm is covered even if they download the model personally for work. Those users need a separate written commercial license from Nace. The revenue and funding limits do not apply to personal use or research.
Who cannot use it. The license does not cover any use, for any purpose, by a person whose employer or affiliated organization offers a product or service that competes with Nace's own, including decision models and document intelligence services. Unlike the revenue and funding limits, this exclusion has no exception for personal use or research.
How it may be used. The license prohibits fully automated decisions with significant effects on people, such as employment, credit or healthcare decisions, without meaningful human review. It also prohibits giving medical, legal or financial advice without qualified human oversight, and uses that discriminate against protected groups.
In practice, Drex Max fits an assistant role: it sorts and flags, and a person decides. This summary is not legal advice. Read the full license before deploying it.
Use probabilities carefully
A probability of 0.99 means the model strongly favors an answer, not that the answer is right. Before relying on Drex Max, take 50 to 100 real examples that your team has already labelled. Run them through the model and count the mistakes, especially among the high-confidence answers. Hold back labelled examples when choosing a threshold. On that held-back set, record how many tickets would be routed automatically and how many of those routes would be wrong. Review mistakes even when the model assigned them a high probability.
Keeping it local
On your own Mac, through MLX, Kapitsa, llama.cpp or a locally built Ollama, the text you send stays on your computer. The same model is also offered through Nace's API, DeepInfra and OpenRouter, and text sent to those services is processed on their servers. If an app offers "Drex Max", check which version it calls before sending client or customer material. The local AI privacy guide lists what to check.
Related reading
- Ollama decision models: sort tickets locally with Nimble
- Model compression (quantization) explained
- RAM and VRAM calculator