Ollama 0.19 Preview: MLX Integration for Apple Silicon Acceleration

Ollama has released a preview version of Ollama 0.19, integrating Apple's MLX machine learning framework to provide the fastest possible performance for running models on Apple Silicon. This update focuses on accelerating demanding local AI workloads, such as coding agents and personal assistants, by leveraging Apple's unified memory architecture and new GPU Neural Accelerators.

MLX-Powered Performance on Apple Silicon

Ollama now utilizes the MLX framework to optimize how models interact with Apple Silicon's unified memory architecture. This integration provides a large speedup across all Apple Silicon devices, specifically targeting improvements in both time to first token (TTFT) and generation speed (tokens per second).

On M5, M5 Pro, and M5 Max chips, Ollama leverages GPU Neural Accelerators to further enhance performance. Based on testing conducted on March 29, 2026, using the Qwen3.5-35B-A3B model, Ollama 0.19 is projected to achieve 1851 token/s prefill and 134 token/s decode when utilizing int4 quantization.

NVFP4 Support and Production Parity

Ollama now supports NVIDIA's NVFP4 format, which reduces memory bandwidth and storage requirements for inference workloads while maintaining model accuracy. This support provides two primary benefits:

1. Production Parity: Users can achieve results consistent with those from large-scale inference providers who use the NVFP4 format.
2. Optimization Compatibility: Ollama can now run models optimized by NVIDIA's model optimizer.

Other precision formats will be made available in the future based on research and hardware partner requirements.

Enhanced Caching for Agentic Workflows

Ollama has upgraded its caching mechanism to improve responsiveness for coding and agentic tasks. These improvements include:

• Cross-Conversation Cache Reuse: Ollama reuses its cache across different conversations, reducing memory utilization and increasing cache hits when branching from a shared system prompt (e.g., when using tools like Claude Code).
• Intelligent Checkpoints: The system now stores cache snapshots at strategic locations within the prompt, which reduces the amount of prompt processing required and accelerates response times.
• Smarter Eviction: Shared prefixes are preserved longer, even when older branches are dropped.

Implementation and Model Support

To utilize these features, users require a Mac with more than 32GB of unified memory. The preview release specifically accelerates the Qwen3.5-35B-A3B model, with sampling parameters tuned for coding tasks.

Users can launch the model for specific tools using the following commands:

• Claude Code: ollama launch claude --model qwen3.5:35b-a3b-coding-nvfp4
• OpenClaw: ollama launch openclaw --model qwen3.5:35b-a3b-coding-nvfp4
• General Chat: ollama run qwen3.5:35b-a3b-coding-nvfp4

Ollama is expanding the list of supported architectures and plans to introduce a simplified import process for users with custom models fine-tuned on supported architectures.