Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon
Nace AI has released Drex Max v1 as open weights. Drex Max is its most capable decision model, and it now runs on your own hardware. The release ships bf16 weights, a Q8_0 GGUF, an affine Q4 checkpoint, and forks of llama.cpp and Ollama. For developers, Drex Max open weights mean routing, classification, and policy checks can stay on a local machine.
TL;DR
- Size: Dense 27B parameters, 262,144-token context window (16,384 by default), and up to 255 options per question.
- Runs on: NVIDIA GPUs with Python (H100 bf16, or L40S 48 GB in 8-bit); Apple Silicon Macs with 24 GB+ (affine Q4) or 48 GB (Q8_0 GGUF).
In 2026, local-first AI deployments have moved from hobbyist experiments to production defaults for many teams. With Drex Max v1, Nace AI positions its decision model squarely in that shift: open weights, Apple Silicon support, and a quantization stack that makes a 27B model viable on a laptop.
What's Included in the Release
The Drex Max v1 release includes:
- bf16 weights for full-precision inference and fine-tuning.
- Q8_0 GGUF for high-fidelity quantized inference.
- Affine Q4 checkpoint for memory-constrained devices.
- Forks of llama.cpp and Ollama with upstream support for the model.
Why a Decision Model?
Unlike general-purpose chat models, Drex Max is optimized for constrained-choice tasks: routing requests between tools or services, classifying inputs, and enforcing policy checks. These workloads typically demand low latency, deterministic output structure, and predictable memory footprints—all of which favor local execution over remote API calls.
The 262,144-token context window (16,384 by default) allows the model to reason over long documents or extended agent histories, while support for up to 255 options per question makes it suitable for high-cardinality classification and decision trees.
Hardware Requirements
| Platform | Configuration |
|---|---|
| NVIDIA GPU | H100 in bf16, or L40S 48 GB in 8-bit, with Python |
| Apple Silicon Mac | 24 GB+ unified memory (affine Q4) or 48 GB (Q8_0 GGUF) |
The Apple Silicon support in particular aligns with a broader 2026 trend: developers increasingly expect capable models to run natively on M-series hardware without cloud dependencies. With 24 GB of unified memory, the affine Q4 checkpoint makes Drex Max accessible on MacBook Pro and Mac Studio configurations that are now standard for professional development.
Getting Started
Developers can pull the model from Hugging Face and use the Nace AI forks of llama.cpp or Ollama for local inference. The availability of both Q8_0 GGUF and affine Q4 checkpoints gives teams a straightforward trade-off between fidelity and memory use, while the bf16 weights remain the reference for evaluation and customization.
Bottom Line
Drex Max v1 is a focused release: a dense 27B decision model with open weights, practical quantization options, and first-class support for local hardware—including Apple Silicon. For teams building agentic systems, routing layers, or policy engines in 2026, it offers a credible path to keeping decision logic on-device.
via MarkTechPost