Ollama MLX Apple Silicon integration is now available in preview with the release of Ollama 0.19, marking a major leap in local AI performance on macOS. Powered by Apple’s MLX machine learning framework, this update takes full advantage of the unified memory architecture found in Apple Silicon chips to deliver dramatically faster inference speeds.
What Is Ollama MLX Apple Silicon and Why It Matters
Until now, Ollama relied on a different backend for macOS inference. With version 0.19, the platform has been rebuilt on top of MLX, Apple’s own open-source machine learning framework designed specifically for Apple Silicon. The result is a significant performance boost across all Apple Silicon devices — from the M1 to the latest M5 series. For developers running AI workloads locally, this is a game-changing update.
Benchmark Numbers: Prefill and Decode Speed
The performance gains are substantial. Testing conducted on March 29, 2026, using Alibaba’s Qwen3.5-35B-A3B model showed the following results on Apple’s M5 chip family:
- Prefill speed: 1,810 tokens/s with Ollama 0.19 vs. 1,154 tokens/s with Ollama 0.18 — a 57% improvement
- Decode speed: 112 tokens/s with Ollama 0.19 vs. 58 tokens/s with Ollama 0.18 — nearly double the output speed
- With int4 quantization: Ollama 0.19 reaches 1,851 tokens/s prefill and 134 tokens/s decode
On M5, M5 Pro, and M5 Max chips specifically, Ollama now leverages the new GPU Neural Accelerators to boost both time to first token (TTFT) and generation speed. This makes Mac machines genuinely competitive for serious local AI inference tasks.
NVFP4 Support and Smarter Caching
Ollama 0.19 also introduces NVFP4 quantization support, a format developed by NVIDIA that maintains model accuracy while reducing memory bandwidth and storage requirements. This brings Ollama users closer to production-level inference parity — meaning the results you get locally will more closely match what cloud inference providers deliver at scale.
Alongside NVFP4, the caching system has been overhauled with three key improvements relevant to coding agents and AI applications:
- Lower memory utilization: The cache is now reused across conversations, reducing memory overhead when using shared system prompts with tools like Claude Code
- Intelligent checkpoints: Snapshots are stored at smart locations in the prompt, reducing redundant processing and speeding up responses
- Smarter eviction: Shared prefixes survive longer even when older conversation branches are dropped
How to Get Started with Ollama 0.19
This preview release is optimized for the Qwen3.5-35B-A3B model with sampling parameters tuned for coding tasks. Users need a Mac with more than 32GB of unified memory. Launch commands are straightforward:
- Claude Code:
ollama launch claude --model qwen3.5:35b-a3b-coding-nvfp4 - OpenClaw:
ollama launch openclaw --model qwen3.5:35b-a3b-coding-nvfp4 - Direct chat:
ollama run qwen3.5:35b-a3b-coding-nvfp4
The Ollama team has acknowledged contributions from the MLX contributor team, NVIDIA contributors to NVFP4 quantization and MLX CUDA support, the GGML and llama.cpp community, and the Alibaba Qwen team for open-sourcing their models. Future updates will expand supported model architectures and introduce easier import workflows for custom fine-tuned models.
