Editorial illustration for Nace AI Open-Sources 27B Drex Max v1 to Run Locally on Apple Silicon
Nace AI Open-Sources 27B Drex Max for Apple Silicon
Nace AI has put its biggest decision model out in the open. Drex Max v1, a dense 27-billion-parameter system, is now downloadable as bf16 weights, a Q8_0 GGUF, and an affine Q4 checkpoint, with forks of llama.cpp and Ollama to run it. That matters because this isn't a chatbot.
Drex Max takes a state (text, JSON, or images) plus a set of named questions, and returns a probability for each option in a single forward pass. No text generation involved. It handles three question types: multiple choice among named options, yes/no, and ordinal scoring.
That structure makes it suited to agent routing, ticket triage, and guardrail checks, the kind of work usually handed off to a hosted API.
With this release, developers can keep those routing and policy decisions on their own hardware instead. The model runs on NVIDIA GPUs through Python, on Apple Silicon Macs with 24GB or more of memory, and on CPU through llama.cpp and Ollama. Nace.AI frames it as a 262,144-token context model built on a 64-layer backbone, and the company has published self-reported benchmark numbers to back up the claim that this is its most capable decision model yet.
Quantization holds up in Nace’s tests. On a 48 GB Apple M5 Pro, the Q8_0 GGUF matched H100 bf16 within 0.02 per probability. The affine Q4 stayed within 0.0329 of Q8_0, with no choice flips.
Why this matters
For teams building agents or routing pipelines, the appeal here isn't the 27B parameter count, it's the quantization math. A 0.02 probability delta between bf16 on an H100 and Q8_0 on a Mac Studio is small enough to stop worrying about, and zero choice flips at Q4 on 48GB of unified memory means decision logic can run on a laptop without second-guessing every output against a cloud baseline. That's the actual unlock for founders who don't want policy checks or classification calls leaving their infrastructure, especially anyone in healthcare, finance, or legal workflows where sending prompts to an external API is a non-starter.
We'd still want independent benchmarks beyond Nace AI's own Decision Index before treating this as settled, since the comparison numbers cited so far come from the vendor's own test suite. The llama.cpp and Ollama forks matter more than the headline size: they're what determine whether this actually slots into existing local-inference setups, or just works in Nace's own demo environment. Worth watching whether third parties replicate the Q4 stability claim.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Factory Droid 1.0 Uses Script Mode to Cut Unit Test Model Calls to 5
From Archives Vercel Labs open-sources agent-desktop for model-agnostic computer use
Common Questions Answered
What is Drex Max v1 and how does it differ from traditional chatbots?
Drex Max v1 is a 27-billion-parameter decision model developed by Nace AI that takes a state (text, JSON, or images) plus a set of named questions and returns a probability for each option in a single forward pass. Unlike chatbots, it does not generate text but instead specializes in decision-making tasks including multiple choice, ranking, and classification questions.
What quantization formats are available for running Drex Max v1 locally?
Drex Max v1 is available in three quantization formats: bf16 weights for full precision, Q8_0 GGUF for moderate compression, and affine Q4 checkpoint for maximum compression. These formats can be run using forks of llama.cpp and Ollama, enabling local deployment on consumer hardware like Apple Silicon machines.
How accurately does the Q8_0 GGUF quantization perform compared to the original H100 bf16 model?
According to Nace AI's testing on a 48 GB Apple M5 Pro, the Q8_0 GGUF matched the H100 bf16 baseline within 0.02 probability difference per choice. This minimal delta means teams can confidently run the quantized version locally without worrying about accuracy degradation compared to cloud-based inference.
What are the practical benefits of running Drex Max v1 on Apple Silicon instead of cloud infrastructure?
Running Drex Max v1 locally on Apple Silicon enables decision logic to execute on laptops without requiring cloud baseline verification or policy checks. The affine Q4 quantization achieves zero choice flips on 48GB unified memory systems, allowing founders to build agents and routing pipelines with full confidence while eliminating latency and dependency on external services.
What types of decision-making questions can Drex Max v1 handle?
Drex Max v1 is designed to handle three question types: multiple choice, ranking, and classification tasks. For each question type, the model processes the input state and returns probabilities for each available option in a single forward pass, making it suitable for decision routing and agent systems.
Further Reading
- Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon - MarkTechPost
- Nace Releases Drex Max Weights; Firms Over $1M Need a Deal - FourWeekMBA
- Drex Max: Nace's Hosted Multimodal Decision Model Reads Up to Eight Images per Request - ModelSystem.One
- Nace Opens Drex Max v1, a 27B Multimodal Decision Model That Runs on a Mac - ModelSystem.One
