27B Model Runs on Apple Silicon, Nace AI Open-Sources Drex Max v1

Nace AI has officially open-sourced Drex Max v1. The MarkTechPost report presents information directly: it is a 27B parameter multimodal decision-making model with a public score of 67.52, and capable of running local inference on Apple Silicon. The original article is written by Asif Razzaq, with the full link placed at the end of this piece.

First let's list all confirmed facts clearly:

Model: Drex Max v1
Developer: Nace AI
Parameter count: 27B
Type: Multimodal decision-making model
Public benchmark score: 67.52
Deployment: Local inference on Apple Silicon
Source: MarkTechPost

For developers, running a 27B model locally comes with several real practical challenges. Apple Silicon's unified memory architecture eliminates the requirement for large models to be hosted on datacenter GPUs by default, but whether the model can run, how fast it runs, and how long its supported context window is all depend on the quantization scheme, memory footprint, inference framework and batching strategy. These details are not specified in the available materials. What is confirmed now is the clear industry direction: large models are steadily moving to local end-user devices.

For general users, the value of local inference is even more straightforward. All user data stays on their own machine, and the cost of using the model shifts from cloud service bills to only electricity consumption and waiting time. The catch here is that local models almost always require tradeoffs between output performance, inference speed, and hardware budget. A model that "can run" still has a notable engineering gap before it becomes reliably usable for daily routines.

How to interpret the 67.52 score? Without the full benchmark name, reference comparison models, and evaluation methodology, the standalone number cannot be used to judge the model's actual performance level. It acts more like a pending reproduction record, rather than a definitive conclusion. For an open-source model to be widely adopted, it also requires complete licensing documents, public full weights, functional inference code, pre-compiled quantized versions, and well-structured documentation.

Therefore, there are two notable points about this open-source release. First, the fact that a 27B multimodal model takes local Apple Silicon as a target environment proves that the boundary of end-side on-device inference is still expanding outward. Second, the score is published alongside the open-source assets, and what comes next will be whether the community can reproduce the reported results, and run the model with acceptable latency on their own local machines.

Original Article:
Asif Razzaq, MarkTechPost
https://www.marktechpost.com/2026/11/01/nace-ai-open-sources-drex-max-v1-local-inference/