---
title: "Ollama's highest performance on Apple Silicon yet with MLX · Ollama Blog"
url: https://daily.dev/posts/ollama-s-highest-performance-on-apple-silicon-yet-with-mlx-ollama-blog-tlj9dedny
source_url: https://ollama.com/blog/mlx-performance
type: article
source: "Ollama"
published: 2026-08-23T12:23:55.975Z
updated: 2026-08-23T12:59:33.334Z
tags: ["data-science", "ai-inference", "ollama"]
reading_time: 3
upvotes: 0
comments: 0
language: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Ollama's highest performance on Apple Silicon yet with MLX · Ollama Blog

**[Ollama](https://daily.dev/sources/ollama)** · 3 min read · 0 upvotes · 0 comments

## Summary

Ollama's MLX engine on Apple Silicon has been updated with several performance improvements: support for NVIDIA's NVFP4 quantization format for higher-quality low-precision outputs, up to 20% faster inference from fused Metal kernels and reworked GPU sampling, and a new snapshot system that caches model state at key points to speed up agentic workflows involving multi-agent handoffs, thinking models, and branching conversations. Users can try it by downloading the latest Ollama and running an MLX-tagged model.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://ollama.com/blog/mlx-performance>

## Questions this post answers

### What is NVFP4 and how does it improve quantized model quality in Ollama's MLX engine?

NVFP4 is NVIDIA's model-optimized 4-bit quantization format that tracks the local dynamic range of model weights more closely than standard formats, reducing quantization loss. In Ollama's MLX engine, it roughly halves the perplexity loss compared to q4_K_M quantization for the Gemma 4 12B model while maintaining similar performance, and also allows datacenter-optimized models to run on desktop hardware.

_daily.dev surfaces updates like this for developers weighing quantization formats for local inference._

### How much faster is Ollama's MLX engine after the latest update?

Ollama's MLX engine is now up to 20% faster due to new optimizations, including fusing several operations into single Metal kernels via MLX's just-in-time compiler and reworking GPU-backed sampling to run more efficiently. These changes target inference speed on Apple Silicon specifically.

_Track engine performance changes like this on daily.dev before choosing a local inference setup._

### How does Ollama's snapshot system speed up agent workflows with repeated tool calls?

Ollama's snapshot system saves model state at key points in a conversation, such as before branching, at intervals through long prompts, and just before each response, so repeated context like system prompts and tool definitions only needs processing once. This helps multi-agent handoffs, thinking models that drop reasoning tokens, and branching or retried responses avoid reprocessing the entire transcript each time.

_Developers building agent workflows can follow engine-level changes like this on daily.dev._

---

Tags: [#data-science](https://daily.dev/tags/data-science), [#ai-inference](https://daily.dev/tags/ai-inference), [#ollama](https://daily.dev/tags/ollama)

[View this post on daily.dev](https://daily.dev/posts/ollama-s-highest-performance-on-apple-silicon-yet-with-mlx-ollama-blog-tlj9dedny)