Most of the open-weights conversation in AI is about models that generate things — images, video, text, music. Nace AI’s Drex Max v1 does something less flashy and arguably more load-bearing for anyone building an actual application: it decides between things. Routing a request to the right handler, classifying content, running a policy check before an action fires — the unglamorous plumbing that sits in front of the generative part of a pipeline. Nace has now released its weights openly rather than keeping Drex Max behind its hosted API alone.
What it actually does
Drex Max is a dense 27B parameter multimodal model with a 262,144-token context window (16,384 by default) built specifically for picking between up to 255 options per question — structured decision-making rather than open-ended generation. The release ships bf16 weights, a Q8_0 GGUF quantization, an affine Q4 checkpoint, and Nace’s own forks of llama.cpp and Ollama for running it. On hardware: it targets an NVIDIA H100 in bf16 or an L40S 48GB card in 8-bit, but the headline claim is running entirely offline on an Apple Silicon Mac with 24GB+ of RAM at Q4, or 48GB for the full Q8_0 GGUF — no cloud round-trip required for routing and classification work.
On Nace’s own self-reported benchmark — the public Decision Index 0.4 — Drex Max v1 scores 67.52, which the company says is 6.27 points ahead of the previous Drex v2.0. It reports a strong 97.3 on InfographicVQA specifically, suggesting the model is tuned toward reading structured visual content (forms, dashboards, diagrams) as part of making a decision, not just plain text.
The license is the actual news
Multiple outlets initially described this as a straightforward open-source release, but the weights ship under Nace’s own Open RAIL-M license (modified), not a permissive license like MIT or Apache-2.0. It’s free to use for individuals, research, and organizations under $1 million in revenue or funding; above that threshold, a separate commercial agreement applies, and the license carries explicit restrictions against using the model for medical, legal or financial decisions without qualified human oversight. That’s a meaningfully different proposition than “it’s open source now” — more like “you can self-host it below a revenue ceiling,” which matters if you’re planning to build on it commercially rather than just experiment.
Why this matters for a creative pipeline
A model purpose-built for routing and classification is exactly the kind of component that sits upstream of a generative workflow — deciding which of several image models to call, flagging content before it reaches a generation step, triaging a batch of user uploads. Having that piece runnable locally on consumer-tier Apple Silicon, rather than requiring a hosted API call per decision, is the part worth tracking even if Drex Max itself never generates a single pixel.