Grux AI Download

Should I run MLX or GGUF models on my Mac?

GGUF and MLX are not rivals so much as two engines with two kinds of model file, and an Apple silicon Mac can run either. GGUF is the file format of llama.cpp (MIT), which uses Metal on a Mac and is the engine Ollama builds on. MLX is Apple's machine learning framework, and Apple's MLX LM package (MIT) runs and fine-tunes MLX models on Apple silicon, pulling them from the MLX Community on Hugging Face. LM Studio runs both: GGUF through llama.cpp, and MLX on Apple silicon. Ollama added an MLX runner on Apple silicon in preview with version 0.19. So pick the app first: Ollama for one command and the biggest library, LM Studio to try both formats side by side, MLX LM to fine-tune. This page quotes one vendor test and has benchmarked neither format itself.

Download Grux AI 3.3.1, free, for macOS 14 or later on Apple silicon.

Grux AI 3.3.1 · sources read 2026-11-18 · published 2026-11-19

Which engine runs which file

Engine or appModel filesWhere it runsLicence
llama.cppGGUFMac, with Metal on Apple silicon, and other platformsMIT
MLX LMMLX models, from the MLX Community on Hugging FaceApple siliconMIT
LM StudioGGUF through llama.cpp, and MLXMac, Windows and Linux; MLX on Apple silicon MacsClosed source, free for home and work
OllamaModels from its own library, run on llama.cpp; an MLX runner in previewMac, Windows, Linux; the MLX runner on Apple siliconMIT

What each engine is good for

What speed difference to expect

Search results are full of MLX against llama.cpp benchmarks, and they disagree with each other, because the result depends on the model, the quantization, the chip and the version. The one test we can quote is Ollama's own: on 29 March 2026, with one Qwen model, Ollama 0.19 read the prompt at 1810 tokens a second against 1154 for 0.18, and generated at 112 against 58. Ollama says it ran the new build on an NVFP4 file and the old one on Q4_K_M. That compares two Ollama builds, not MLX against llama.cpp, and it is one vendor's test.

What we have measured is Grux Local on Ollama's MLX runner on a 16 GB Mac mini, on how fast is a local LLM on a Mac.

Where to check this

ClaimSource
llama.cpp is MIT, runs GGUF, and uses Metal on Apple siliconllama.cpp README
MLX LM runs and fine-tunes models on Apple silicon, from the MLX CommunityMLX LM README
LM Studio runs llama.cpp (GGUF) and, on Apple silicon Macs, MLXLM Studio documentation
Ollama is powered by MLX on Apple silicon in preview, from 0.19; the test figuresOllama blog, Ollama is now powered by MLX
Ollama builds on llama.cppOllama README, supported backends

Questions

Is MLX faster than GGUF on a Mac?
It depends on the model, the quantization and the chip, and published tests disagree. Ollama's own test shows its MLX build faster than its previous build, which is not MLX against llama.cpp.
Can LM Studio run both MLX and GGUF?
Yes. LM Studio runs GGUF models through llama.cpp, and on Apple silicon Macs it also runs MLX models.
Does Ollama use MLX?
On Apple silicon, in preview, since version 0.19. Ollama still builds on llama.cpp for its other models.
Which should a beginner use?
Ollama, for one command and the largest library. Move to LM Studio when you want to compare formats, or MLX LM when you want to fine-tune.
Download Grux AI 3.3.1 Read the source

Free, MIT licensed. macOS 14 or later, Apple silicon. 29.1 MB, signed and notarized by Apple. No account, no server, no subscription.