Ollama is now powered by MLX on Apple Silicon in preview
ollama.com
March 31, 2026
3 min read
🔥🔥🔥🔥🔥
70/100
Summary
Ollama is now powered by MLX on Apple Silicon, offering significantly improved performance for applications on macOS. This enhancement accelerates personal assistants like OpenClaw and coding agents such as Claude Code and OpenCode.
Key Takeaways
- Ollama is now powered by Apple's MLX machine learning framework, resulting in significantly improved performance on Apple Silicon devices.
- The integration of NVIDIA’s NVFP4 format allows Ollama to maintain model accuracy while reducing memory bandwidth and storage requirements for inference workloads.
- Ollama's cache has been upgraded to enhance efficiency, resulting in lower memory utilization and faster response times for coding and agentic tasks.
- The preview release accelerates the Qwen3.5-35B-A3B model, optimized for coding tasks, and requires a Mac with more than 32GB of unified memory.
Community Sentiment
Positives
- Running LLMs on device enhances security and addresses the demand for inference, potentially reducing electricity consumption, which is crucial for sustainable AI deployment.
- The transition to native MLX on Apple Silicon is expected to improve memory handling, which could lead to better performance for larger models.
Concerns
- There is uncertainty about the performance of Ollama compared to other models and inference engines, indicating potential limitations in its capabilities.
- Users express frustration about the current hardware limitations, such as running advanced models on systems with only 16GB of RAM.
