How to Run Drex Max v1 on Mac: Nace AI's Open-Weight 27B Decision Model Explained

how to run Drex Max v1 on Mac — a MacBook Pro laptop, illustrative image

Illustrative image: a MacBook Pro. Drex Max v1 is software, not a device — it runs on Apple Silicon Macs like this one. Image: Wikimedia Commons.

TL;DR — How to run Drex Max v1 on Mac, in one line: download Nace AI's open weights from Hugging Face, install Kapitsa or the Nace Ollama fork, and serve the 27-billion-parameter decision model on your own machine — no cloud API needed for text decisions.

📌 KEY HIGHLIGHTS

Open weights, released 1 November 2026: Nace AI has published Drex Max v1 under its Open RAIL-M licence, after the model itself launched as a hosted API in October.
It decides, it does not chat: you send a state plus named questions and get back a probability for every option in one forward pass — no generated text.
Mac-friendly builds: a 4-bit checkpoint runs on Apple Silicon Macs with 24 GB of memory or more; the 8-bit build needs 48 GB.
Local means private: routing, classification and policy checks can stay on your machine instead of going to a hosted API.
Text only, locally: the Mac and CPU builds read text — image input still needs the Python weights on a GPU.

How to run Drex Max v1 on Mac: what you need first

Drex Max is a decision model, not a chatbot. A normal language model writes an answer token by token; Drex Max takes a situation (text, JSON or images) plus questions you define — pick one of these options, yes or no, or a score on a scale — and returns a calibrated probability for each option. That makes it a fit for agent routing (which tool should run next?), ticket triage and guardrail checks, where you want a number you can threshold, not a paragraph.

Before you start, check your Mac against the table below. The headline numbers come from Nace AI's own model card on Hugging Face, so treat the benchmark scores as company-reported, not independently verified.

Build	Size	Mac you need	Images?
Affine Q4 (Kapitsa / MLX)	14.45 GB	Apple Silicon, 24 GB+ memory	No — text only
Q8_0 GGUF (llama.cpp / Ollama / Kapitsa)	27.25 GB	Apple Silicon, 48 GB memory	No — text only
bf16 Python weights	~49 GB	NVIDIA GPU (H100 class)	Yes

Step-by-step: run Drex Max v1 on your Mac

Step 1 — Install the Hugging Face CLI. You need it to download the weights. In Terminal: pip install -U huggingface_hub.

Step 2 — Download the weights. For the 8-bit build: hf download nace-ai/drex-max-v1-Q8_0 drex-max-v1-Q8_0.gguf --local-dir weights/drex-max-v1-Q8_0. On a 24 GB Mac, fetch the affine Q4 checkpoint instead and use Kapitsa or MLX.

Step 3 — Pick the easiest runtime: Kapitsa. Kapitsa v0.2.25 or later imports the GGUF and serves it. Install it with brew install nace-ai/tap/kapitsa, then import and serve: kapitsa import weights/drex-max-v1-Q8_0/drex-max-v1-Q8_0.gguf --name drex-max-v1 followed by kapitsa-systemone --model "$HOME/.cache/kapitsa/models/drex-max-v1" --port 8000 --context 2048 (on a 48 GB Mac, Nace recommends context 2048).

Step 4 — Or use the Ollama fork. Use Nace's fork of Ollama (branch drex-max-v1), place this Modelfile next to the GGUF, and note the CAPABILITY decision line — without it, Ollama treats the file as a chat model:

FROM ./drex-max-v1-Q8_0.gguf
CAPABILITY decision
PARAMETER num_ctx 16384
PARAMETER num_gpu 40

Then run ollama create drex-max-v1 -f Modelfile. A third option is the llama.cpp fork: build its drex-max-v1 branch and start llama-server with -np 1 -c 16384 -b 16384 -ub 16384 --no-warmup --port 8000.

how to run Drex Max v1 on Mac — a laptop computer, illustrative image

Illustrative image: a laptop. The local builds serve the same API as the hosted model. Image: Wikimedia Commons / KK IN HK, CC BY-SA.

Step 5 — Send your first decision request. Every runtime serves the same POST /v1/systemone endpoint. A minimal test against Ollama looks like this:

curl localhost:11434/v1/systemone -d '{"model": "drex-max-v1", "state": "...", "questions": {...}}'

Replace the state with your situation and the questions with your options. The response gives a probability for every option — pick the highest, or set a threshold below which your agent asks a human.

Step 6 — Know the accuracy trade-offs. Nace reports that on a 48 GB Apple M5 Pro, the Q8_0 build matches the H100 bf16 weights within 0.02 on every probability, and the 4-bit build stays within 0.0329 of Q8_0 with no choice flips. In plain terms: quantisation barely moves the answers. But remember the local builds are text-only — if your agent needs screenshot or image decisions, you need the Python GPU weights.

Why it matters for World Tech readers

Decision models are a quieter branch of the AI tree than chatbots, but they are the part that lets agents act safely: a coding agent that must choose between three tools, or a support system that must route a ticket, needs a reliable probability, not fluent prose. Until now, Nace's strongest model in this family was hosted-only — you sent your data to its API. Open weights change who can use it: developers in India and elsewhere can now run the same decision engine on a Mac, keep customer text on-device, and avoid per-request API costs and latency. The benchmark headline — 67.52 on the public Decision Index 0.4, against 61.25 for Drex v2.0 — is Nace's own figure, and the model trails rivals on some knowledge tests, so treat it as a specialist tool, not a general brain.

Should you care?

✅ Verdict: Yes, if you build agents. If you have a 24 GB+ Apple Silicon Mac, the 4-bit build is a free, private way to add calibrated decisions to your projects today. If you only want a chatbot, this is not it — Drex Max writes no text at all. Watch for an image-capable local build, which is the missing piece.

FAQs

What is a decision model?
A model that answers questions with probabilities instead of text. You give it a state and typed questions — choice, yes/no or a score — and one forward pass returns the probability of every option. It is built for routing, triage and checks inside agents.

What Mac do I need to run Drex Max v1?
An Apple Silicon Mac with at least 24 GB of unified memory for the affine Q4 checkpoint (via Kapitsa or MLX), or 48 GB for the Q8_0 GGUF (via llama.cpp, Ollama or Kapitsa). On a 48 GB Mac, Nace recommends context 2048 for Kapitsa.

Can Drex Max see images when run locally?
Not yet. The local GGUF builds cover the text backbone only. Image input needs the full Python bf16 weights on an NVIDIA GPU. Nace's hosted API is the only path that reads images today.

Is Drex Max free to use?
The weights are published under the Nace.AI Open RAIL-M licence, which allows use and redistribution with conditions — read the licence on the model page before shipping a commercial product. The hosted API remains a separate, paid option.

Takeaway: Drex Max v1's open weights turn a hosted decision engine into something you can run on a Mac: download, serve, and get probabilities instead of prose.

Follow World Tech on TechByCountry for tomorrow's update.

📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.

TechByCountry — Daily Tech News, Country by Country

Search This Blog