Artificial intelligence developer Nace AI has released Drex Max v1 as an open-source project, publishing a 27-billion-parameter multimodal decision model configured to run locally on Apple Silicon hardware. The release marks an expansion of on-device systems capable of handling visual and textual inputs without relying on external cloud servers.
Key points
- Nace AI has published the weights for its 27-billion-parameter multimodal decision architecture.
- The system achieved an evaluation score of 67.52 on benchmark testing reported by the project.
- The architecture runs locally on Apple Silicon chips by taking advantage of unified system memory.
- The release reflects growing demand for private, on-device multimodal inference on personal workstations.
Capabilities and architecture
The system functions as a multimodal decision model rather than a conventional text-generating language model. At 27 billion parameters, the architecture is sized to balance contextual evaluation across modalities while remaining small enough to load into workstation memory.
According to technical specifications published alongside the release, the model recorded an evaluation score of 67.52. Nace AI published the weights under an open-source licence, allowing researchers and software engineers to inspect the architecture, run local tests, and integrate the system into desktop workflows.
Multimodal decision models differ from conversational assistants by prioritising planning, discrete action selection, and reasoning across structured visual and text inputs. Deploying these architectures at 27 billion parameters allows automated agents to process rich input streams while assessing multi-step tasks.
Running large models on Apple Silicon
A central feature of the release is native compatibility with Apple Silicon processors. Running a 27-billion-parameter network locally has typically required high-end desktop graphics cards with substantial dedicated video memory.
Apple’s M-series chips use a unified memory architecture, allowing the central processor, graphics cores, and neural engine to draw from a single memory pool. Mac computers equipped with 32 gigabytes, 64 gigabytes, or higher capacities can therefore hold the model weights in memory without requiring multi-card server setups.
Local execution eliminates the network latency common in remote cloud services and ensures data remains on the physical machine. For developers working under strict privacy constraints, on-device execution prevents exposure to shared cloud providers.
Open weights and local software distribution
The decision to publish the weights freely arrives amid broader shifts in how open machine learning assets are distributed and hosted. While commercial frontier laboratories often restrict access through paid application interfaces, independent research groups increasingly target consumer-accessible hardware.
Developers working with open ecosystems have faced scrutiny over repository security and distribution channels, as seen when earlier investigations into download platforms highlighted vulnerabilities like the Hugging Face malware campaign that distributed disguised software. As open releases expand into decision-making agents, verifying local model weights has become a standard requirement for engineering teams.
Regulators have also intensified their focus on how open models and compute resources intersect, including the wider Nvidia Hugging Face inquiry into software distribution and market practices. Nace AI’s independent release provides an alternative for technical users seeking standalone local deployment without vendor lock-in.
What happens next
Independent evaluations are expected to follow from the open-source community, testing how the 67.52 benchmark score translates across robotics, workflow automation, and multimodal reasoning tasks. Developers can download the model weights to test inference speeds and memory consumption on Apple hardware.
