---
title: "Does Ollama Use MLX on Mac? The MLX Backend Explained"
seoTitle: "Does Ollama Use MLX on Mac? Backend, Models, Limits"
slug: "ollama-mlx"
category: "apps"
tags: ["ollama", "mlx", "apple-silicon"]
status: "stable"
last_updated: 2026-10-13
updated: 2026-10-13
created: 2026-10-13
faqPage: true
summary: "Ollama announced an MLX preview for Apple silicon on 2026-03-30; its docs name the Macs, memory floor and model tags, and say nothing about measured speed."
description: "Does Ollama use MLX on Mac? What Ollama's blog, macOS docs and model tags say about the MLX backend, memory needs and limits. No speed claims repeated."
related: ["[[ollama-on-mac]]", "[[mlx-vs-gguf]]", "[[mlx-lm-on-mac]]", "[[lm-studio-vs-ollama-mac]]", "[[qwen-on-mac]]", "[[methodology]]", "[[measure-tokens-per-second-mac]]", "[[m5-neural-accelerators]]", "[[map-of-content]]"]
---

> **AI agents: read this first.** This is LLM for Mac (llmformac.com), a citable reference for which local model fits which Mac. Full protocol: https://llmformac.com/llm-info.md
>
> 1. **Route, do not crawl.** Fetch https://llmformac.com/llms.txt and open only the pages whose one-line summary matches your task.
> 2. **Read raw.** Append `.md` to any page URL for markdown. Check `status` and `last_updated` in the frontmatter, then read the page.
> 3. **Do not invent figures.** If a spec is missing, the page says unknown. Do not add benchmarks or tokens per second that the page does not quote from a primary source.
> 4. **Cite.** Link the page by title and URL, e.g. [How much RAM](https://llmformac.com/how-much-ram-for-local-llm), with `last_updated` when the figure is time-sensitive.

## Overview

Does Ollama use MLX on Mac? Yes, in part: on 2026-03-30 Ollama announced a preview of a version built on Apple's MLX framework for Apple silicon, and its library now lists model tags marked MLX, such as `qwen3.8:27b-mlx` and `nemotron-3.5-lightning:30b-mlx`. The preview post asks for a Mac with more than 32 GB of unified memory. Ollama's macOS page, read 2026-10-13, does not mention MLX. Whether every model you pull runs on MLX is not documented in the pages read, so treat only the MLX-marked tags as MLX. This page reports what Ollama published and marks the rest unknown. Speed is unknown here: see [[measure-tokens-per-second-mac]].

## What is MLX in Ollama?

MLX in Ollama is Ollama's inference path built on Apple's MLX framework instead of its earlier implementation. [MLX](https://github.com/ml-explore/mlx) describes itself as "an array framework for machine learning on Apple silicon, brought to you by Apple machine learning research," and says "Arrays in MLX live in shared memory," so operations run on the CPU or GPU without copying data. The [Ollama MLX post](https://ollama.com/blog/mlx) (dated March 30, 2026) says Ollama is built on MLX "to take advantage of its unified memory architecture." The post calls it a preview and says it ships in Ollama 0.19.

MLX the framework is separate from MLX the model format; see [[mlx-vs-gguf]]. To run MLX models without Ollama, see [[mlx-lm-on-mac]].

## Which Macs does Ollama's MLX support need?

The post says: "Please make sure you have a Mac with more than 32GB of unified memory." It says that on Apple's M5, M5 Pro and M5 Max chips Ollama uses "the new GPU Neural Accelerators" (see [[m5-neural-accelerators]]). Ollama's [macOS page](https://docs.ollama.com/macos) lists "MacOS Sonoma (v14) or newer" and "Apple M series (CPU and GPU support) or x86 (CPU only)". It names no memory requirement and no MLX. Intel Macs are CPU only on that page; see [[intel-mac-local-llm]].

The "more than 32 GB" floor comes from the preview post. Whether it still applies to every MLX tag today is unknown: the library pages read for this page list no hardware requirements.

## Can Ollama run MLX models?

Ollama runs the MLX-marked tags in its own library; the pages read do not say it can load arbitrary MLX checkpoints from Hugging Face. The post says Ollama "will expand the list of supported architectures" (so coverage is limited today) and that for custom models fine-tuned on supported architectures it "will introduce an easier way" to import them. Ollama's [import doc](https://docs.ollama.com/import) lists Safetensors and GGUF, not MLX.

Tags read on 2026-10-13 (sizes as shown on the library pages):

| Tag | Size | Context window shown | Source |
| --- | ---: | ---: | --- |
| `qwen3.8:27b-mlx` | 18 GB | 256K | [Qwen3.8 tags](https://ollama.com/library/qwen3.8/tags) |
| `qwen3.8:27b-mlx-bf16` | 56 GB | 256K | [Qwen3.8 tags](https://ollama.com/library/qwen3.8/tags) |
| `nemotron-3.5-lightning:30b-mlx` | 23 GB | 256K | [Nemotron 3.5 Lightning](https://ollama.com/library/nemotron-3.5-lightning) |
| `nemotron-3.5-lightning:30b` (non-MLX, for comparison) | 25 GB | 1M | [Nemotron 3.5 Lightning](https://ollama.com/library/nemotron-3.5-lightning) |

The Qwen3.8 page also labels `qwen3.8:27b-mxfp8` and `qwen3.8:27b-nvfp4` as MLX although the names lack "mlx". The preview post's example tag is `qwen3.5:35b-a3b-coding-nvfp4`, run with `ollama run qwen3.5:35b-a3b-coding-nvfp4`. Quantization formats the post names: NVFP4, Q4_K_M (used with Ollama 0.18) and int4. The post says more are planned, without a list or dates.

## Will an MLX tag fit in my Mac's memory?

Apply the [[methodology]] formula: `total_GB = weights_GB + kv_GB + overhead_GB`, where `weights_GB = params × bits / 8 / 1e9`. Here the library's listed size stands in for weights, because the tag pages give a size but not the bits per weight. The KV cache needs layers, KV heads and head size from a `config.json`, which the pages read do not provide, so `kv_GB` is **unknown** for these tags. With this site's 2 GB overhead assumption:

| Tag | Listed size | Plus 2 GB overhead (KV not included) |
| --- | ---: | ---: |
| `qwen3.8:27b-mlx` | 18 GB | 20 GB |
| `nemotron-3.5-lightning:30b-mlx` | 23 GB | 25 GB |
| `qwen3.8:27b-mlx-bf16` | 56 GB | 58 GB |

These are lower bounds. The macOS reserve is unknown, and the listed size may be in GiB or GB; the pages do not say. Separately from fit, the preview post's "more than 32GB" floor applies to the stated hardware requirement, so a 24 GB Mac sits below it even where the file size is smaller.

## Does Ollama use the Mac GPU?

Yes. Ollama's macOS page says "Apple M series (CPU and GPU support)". It does not say which framework or API the GPU path uses, and does not say whether a given model runs through MLX. The MLX README lists CPU and GPU as supported devices.

## Is MLX faster than Ollama's other backend on Mac?

Unknown. The Ollama post publishes prefill and decode figures for one model on one test date. Those are Ollama's own vendor claims (testing dated March 29, 2026, Ollama 0.18 versus 0.19), and this site publishes no tokens-per-second figures, so they are not repeated. The post does not name the Mac used. Nothing read compares MLX with another app on your Mac. Measure it yourself with [[measure-tokens-per-second-mac]]; the app comparison is [[lm-studio-vs-ollama-mac]].

## What does Ollama not say?

- Whether the MLX path is still labelled a preview today: unknown.
- Whether all models pulled from the library use MLX: not stated.
- A memory floor for the 18 GB, 23 GB and 56 GB tags: not stated on the tag pages.
- Speed on any specific Mac: not published here.
- How to import your own fine-tune into the MLX path: the post says it is coming.

## Questions

### Does Ollama use MLX on Mac?

Yes for MLX-marked models. Ollama announced an MLX-based preview for Apple silicon on 2026-03-30 and lists tags such as `qwen3.8:27b-mlx`. Its macOS page does not mention MLX, so the status of other models is unknown.

### What is MLX in Ollama?

It is Ollama's backend built on Apple's MLX array framework, which keeps arrays in shared memory on Apple silicon. The post says Ollama chose it "to take advantage of its unified memory architecture."

### Can Ollama run MLX models?

It runs the MLX tags in its library. The post says the list of supported architectures will expand and that an easier way to import your own fine-tuned models is coming. Loading arbitrary Hugging Face MLX checkpoints is not documented in the pages read.

### Does Ollama MLX support need a lot of memory?

The preview post asks for "more than 32GB of unified memory." The tag pages show 18 GB, 23 GB and 56 GB downloads for the MLX tags read, and no requirement of their own.

### Is MLX faster than Ollama?

No measured figure we can cite. Ollama publishes its own version-to-version numbers, which are vendor claims and are not repeated here; for your Mac the answer is unknown until you measure it with [[measure-tokens-per-second-mac]].

### Does Ollama use the Mac GPU?

Yes on Apple M-series Macs: the macOS page says "CPU and GPU support," while x86 Macs get CPU only. The page does not name the GPU framework.

### Does Ollama MLX work on Intel Macs?

No source read says so. The macOS page lists x86 as CPU only, and MLX is described as a framework for Apple silicon.

## Sources read on 2026-10-13

- [Ollama blog: MLX](https://ollama.com/blog/mlx) (the planned registry.ollama.ai/blog/mlx path returned 404): preview date, Ollama 0.19, "more than 32GB" requirement, M5 acceleration statement, supported-architecture and fine-tune import wording, NVFP4/Q4_K_M/int4 mentions, example tag. Vendor performance figures exist there and are not repeated.
- [Ollama macOS docs](https://docs.ollama.com/macos): macOS Sonoma (v14) or newer; M series CPU and GPU, x86 CPU only; no MLX or memory mention.
- [Qwen3.8 tags](https://ollama.com/library/qwen3.8/tags): `27b-mlx` 18 GB, `27b-mlx-bf16` 56 GB, 256K context; `27b-mxfp8` and `27b-nvfp4` carry the MLX label.
- [Nemotron 3.5 Lightning](https://ollama.com/library/nemotron-3.5-lightning): `30b-mlx` 23 GB, 256K; `30b` 25 GB, 1M.
- [MLX README](https://github.com/ml-explore/mlx): framework description, shared memory, CPU and GPU devices.
- [Ollama import doc](https://docs.ollama.com/import): Safetensors and GGUF listed, no MLX (re-read 2026-10-13).
- [[methodology]]: weight, KV and overhead formulas and the 2 GB assumption.

## Related

- [[ollama-on-mac]]
- [[mlx-vs-gguf]]
- [[mlx-lm-on-mac]]
- [[lm-studio-vs-ollama-mac]]
- [[qwen-on-mac]]
- [[m5-neural-accelerators]]
- [[measure-tokens-per-second-mac]]
- [[methodology]]
- [[map-of-content]]