Ollama’s latest preview release swaps in Apple’s MLX framework as the backend for Apple Silicon Macs, and the speed difference is noticeable. MLX takes advantage of Apple’s unified memory architecture, where the CPU and GPU share the same memory pool, which cuts down on data transfer overhead and improves both latency and throughput.
Right now, MLX support is limited to one model: Qwen3.5-35B-A3B, which is optimized for coding tasks. More models are expected to follow. You’ll need a Mac with more than 32GB of unified memory to run it.
The release also adds a few other things worth noting:
<ul>
<li>NVFP4 quantization (NVIDIA’s low-precision format) is now supported, letting larger models fit into tighter memory budgets without as much accuracy loss as older quantization methods.</li>
<li>Improved KV caching with smarter eviction and intelligent checkpoints, aimed at coding and agentic workloads where context windows tend to get long.</li>
</ul>
The update works with tools like Claude Code and OpenClaw, the latter being one of the local AI agent setups that’s been driving interest in running models on-device. Running locally means your data stays on your machine and you’re not paying per-token API costs — the tradeoff has traditionally been speed, which is exactly what this MLX integration is trying to close.

Collections

Ollama's latest preview release integrates Apple's MLX framework as the backend for Apple Silicon Macs, leveraging unified memory architecture to reduce data transfer overhead and improve inference speed. Currently limited to the Qwen3.5-35B-A3B coding model (requires 32GB+ unified memory), with more models planned. The release also introduces NVFP4 quantization support for fitting larger models into tighter memory budgets, and improved KV caching with smarter eviction for long-context coding and agentic workloads.

Ollama adds MLX backend for Apple Silicon, with faster inference in preview