Running local models on Macs gets faster with Ollama’s MLX support

Ollama, a pioneering runtime platform for executing large language models on personal computers, has announced its integration with Apple's open-source MLX framework for machine learning. This exciting update brings enhancements in caching performance and introduces support for Nvidia’s NVFP4 format, which optimizes model compression and improves memory efficiency across various models. These advancements are particularly beneficial for Macs powered by Apple Silicon chips (M1 or later), as they promise a significant boost in performance. The timing of this development is crucial; local models are gaining popularity beyond academic and hobbyist circles. The recent success of OpenClaw, which soared to over 300,000 stars on GitHub and sparked widespread interest with experiments like Moltbook, has motivated many to explore running models directly on their devices. As developers face challenges such as rate limits and expensive subscriptions for top-tier tools like Claude Code or ChatGPT Codex, the push towards local coding models is intensifying. Additionally, Ollama has recently enhanced its integration with Visual Studio Code, further streamlining the development process. Currently, this new support is available in preview mode with Ollama version 0.19, and it exclusively supports the 35-billion-parameter variant of Alibaba’s Qwen3.5 model. However, potential users should note that the hardware requirements are demanding; a compatible Apple Silicon Mac and at least 32GB of RAM are necessary to utilize these capabilities, as detailed in Ollama’s announcement.