Ollama delivers up to 20% faster Apple Silicon inference: MLX engine, NVFP4 quantization, and snapshots for agentic workflows

Ollama has released the latest update to its MLX engine for Apple Silicon, delivering significant improvements in token generation speed, quantization precision, and support for agentic workflows. The update positions Apple Silicon as a serious alternative to specialized inference hardware for running modern, demanding language models.

What does the MLX engine update mean for Apple Silicon users?

MLX (Machine Learning eXecution) is Apple’s machine learning framework optimized for the Apple Silicon architecture and its unified memory. Ollama’s MLX engine update delivers higher generation speeds, better output quality, and shorter time to first token across all Apple Silicon devices. Beyond the speed gains, the update introduces three key technological additions: NVFP4 quantization, a snapshot system for agentic workflows, and native support for thinking models.

Up to 20 percent faster token generation

The updated MLX engine achieves up to 20 percent faster token generation on Apple Silicon devices. The speed increase is achieved through two main mechanisms:

• Fused Metal kernel operations — combining multiple operations into one reduces communication overhead between processor cores, directly increasing throughput
• GPU-backed sampling — the process of selecting the next token now runs entirely on the GPU, eliminating costly data transfer paths between processor subsystems

The combination of these optimizations means that Ollama users on MacBook Pro, Mac Studio, and Mac Pro devices receive noticeably faster model responses with no configuration changes.

NVFP4: a new quantization tier for desktop inference

One of the most significant additions in this update is support for NVIDIA’s NVFP4 quantization format. Quantization is a technique that reduces the precision of model weights to achieve faster execution and a smaller memory footprint, at an inevitable cost to quality.

NVFP4 strikes an exceptional balance: the format halves the quality loss of 4-bit quantization compared to unquantized BF16 weights. In simpler terms, models quantized with NVFP4 retain significantly more detail and precision than models compressed with standard 4-bit quantization — the difference between a model that stumbles on complex reasoning and one that handles it reliably.

Ollama tested performance on Gemma 4 12B running on a MacBook Pro M5 Max, demonstrating that datacenter-class models are now accessible for desktop and laptop deployment at an acceptable quality level.

Snapshot system for agentic workflows

The new update introduces a snapshot system that directly addresses one of the key challenges in agentic workflows: redundant processing of the same context when branching conversations or retrying steps.

The snapshot system works in several ways:

• Saving model state at conversation branch points and at key processing intervals
• Supporting multiple concurrent agents — each agent can resume work from its own independently saved snapshot, without interfering with others
• Supporting thinking models where reasoning tokens are discarded between turns — snapshots ensure context does not have to be reprocessed from the beginning
• Branching and retry scenarios without the need for full context reprocessing

The snapshot system works alongside prefix caching, which minimizes redundant computation between tool calls and multi-agent handoffs.

Prefix caching and thinking models

Alongside the snapshot system, Ollama introduces native support for thinking models — a category of models that generate an internal chain-of-thought before providing a final answer. Prefix caching reduces the computational cost of repeated contexts that appear in long conversations and multi-agent systems, making agentic applications more economical and faster.

Running Gemma 4 12B on Apple Silicon

All described capabilities are available now. Users can try the updated MLX engine by running Gemma 4 12B via the MLX backend:

ollama run gemma4:12b-mlx

For agentic applications, launching via ollama launch pi --model gemma4:12b-mlx is also available. The update affirms Ollama’s vision: Apple Silicon as a genuine platform for desktop inference of datacenter-class models, leveraging Apple’s unified memory architecture and Metal-based MLX framework for maximum efficiency on consumer hardware.