Every Mac with an Apple Silicon chip can now run powerful artificial intelligence models locally — no internet connection, no subscription, and no data leaving the device.
The update, released in preview as Ollama 0.19, means Mac users can run AI coding assistants, personal chatbots, and productivity tools entirely on their own hardware. The key technical change is the integration of Apple’s MLX machine learning framework, which taps directly into the unified memory shared by the CPU and GPU in every M-series chip.
What MLX changes for Mac users
MLX is Apple’s open-source machine learning framework, purpose-built for the unified memory architecture in Apple Silicon chips. Unlike traditional setups where data moves between the CPU and a separate GPU, MLX allows AI models to access all available memory at once. This eliminates a longstanding bottleneck and makes it practical to run large language models on consumer hardware.
In benchmarks using Alibaba’s Qwen3.5-35B model, Ollama with MLX achieved 1,810 tokens per second during prompt processing — up from 1,154 tokens per second using the previous Metal backend. Token generation speed, which determines how fast the AI responds, increased from 34 to 42 tokens per second.
On Apple’s newest M5, M5 Pro, and M5 Max chips, Ollama goes a step further by leveraging the new GPU Neural Accelerators, boosting both time to first token and sustained generation speed.
Practical uses right now
The practical applications are already here. Ollama 0.19 supports AI coding assistants like Claude Code, OpenCode, and Codex, as well as personal AI agents that handle tasks like summarising documents, drafting emails, or searching local files — all without an internet connection.
For non-technical users, the appeal is simpler: a chatbot that runs entirely on your laptop, with no monthly fee and no data sent to external servers.
The current limitations are worth noting. The preview release supports only one model — Alibaba’s Qwen3.5-35B — and requires at least 32 GB of unified memory. Broader model support and lower memory thresholds are expected in future releases.

Why Australian businesses should pay attention
Australia’s growing focus on data sovereignty makes local AI particularly relevant. Under the Privacy Act 1988 and proposed reforms, Australian businesses face increasing scrutiny over where sensitive data is processed. Running AI models on a local Mac means client information, financial records, and internal documents never leave the device.
For professional services firms — lawyers, accountants, consultants — the ability to use AI without exposing client data to third-party servers addresses a practical compliance concern. It also removes the dependency on stable internet connections, which can be unreliable in regional and rural Australia.
The cost argument is also compelling. ChatGPT Plus costs $30 per month, and enterprise AI subscriptions can run into thousands. A Mac that can run capable AI models locally turns a one-time hardware investment into an ongoing productivity tool with no recurring fees.
A turning point for local AI
The integration of MLX into Ollama marks a shift in how AI computing could work for everyday Mac users. Apple’s hardware has quietly become one of the most capable platforms for running AI models locally, and this update makes that capability significantly more accessible.
The timing aligns with a broader industry trend toward local computing. As cloud AI subscriptions grow more expensive and privacy regulations tighten, the case for on-device processing strengthens. Apple’s unified memory design gives M-series Macs a structural advantage that most Windows and Linux machines cannot easily match.
Whether Apple formalises local AI capabilities in macOS itself remains to be seen. For now, tools like Ollama are filling the gap, and this update makes the barrier to entry lower than it has ever been.
Businesses looking to automate document workflows while keeping data on Australian infrastructure can explore Aivy’s document automation solutions.
