MLX Model Format Explained: Supported Engines and Models

What Is MLX?

MLX is a machine-learning framework for Apple Silicon (M-series chips), developed by Apple’s machine learning research team. Its companion package mlx-lm runs LLMs, and Hugging Face hosts many models converted and quantized for MLX (for example in the mlx-community organization). When we say “MLX format", we mean weights prepared to be loaded by MLX.

Why It Matters

  • It uses the Mac’s unified memory directly. CPU and GPU share one pool of memory on Apple Silicon, so a Mac with plenty of memory can run models too large for a consumer GPU.
  • It is optimized for Apple Silicon. On the same Mac it can generate faster than GGUF (llama.cpp) in some cases; which one wins depends on the model and quantization.
  • 4-bit, 8-bit and other quantized builds are common. A repository name ending in -4bit or -8bit tells you the level.

Tips for Running It Locally

  • Mac (Apple Silicon) only. It does not run on NVIDIA GPUs under Windows or Linux; choose GGUF or AWQ there.
  • LM Studio loads MLX models directly. From the command line, pip install mlx-lm and use mlx_lm.generate or mlx_lm.server.
  • Not all memory is available to the GPU. macOS caps how much memory the GPU can use, so leave headroom rather than picking a model that barely fits your installed memory.
  • The files use the .safetensors extension. Tell MLX builds apart by the repository name or the mlx tag.

Sources: the ml-explore/mlx README, the ml-explore/mlx-lm README and the mlx-community organization on Hugging Face (all as of 2026-09-27).

Our Coverage and Data

Local Model Watch has published 6 article(s) on models available in MLX: 2 where the repository itself is in MLX, and 4 where we found a MLX build of the model. The lists below only include builds we have checked (the publisher’s organization and well-known quantizers); a model missing here may still have a MLX build elsewhere. Part of our model format index.

Main Engines That Load This Format

Engine Overview
exo Framework that clusters everyday devices (Macs, PCs, phones) to run a single model across them.

Models Available in MLX

→ Scroll horizontally to see all columns

Published Model Where to get it Quantizations Smallest VRAM tier Article
2026-09-26 Qwen/Qwen3.8-27B lmstudio-community/Qwen3.8-27B-MLX-4bit, lmstudio-community/Qwen3.8-27B-MLX-8bit MLX 4bit, MLX 8bit 24GB Qwen3.8-27B Multimodal Vision-Language Model: 8GB+ VRAM, GGUF Builds
2026-09-18 prism-ml/Ternary-Bonsai-2-27B-gguf prism-ml/Ternary-Bonsai-2-27B-mlx-2bit MLX 2bit 12GB Ternary-Bonsai-2-27B-gguf Text Generation Model: 8GB+ VRAM
2026-09-17 harshatheg/Qwen-2.5-1B-RLCD This repository — 4GB Qwen-2.5-1B-RLCD Text Generation Model: 4GB+ VRAM
2026-09-11 Edge0/Edge0-35B-A3B-preview This repository — 24GB Edge0-35B-A3B-preview 35B MoE Model for Phone-Class Memory: 24GB+ VRAM
2026-09-09 nex-agi/Nex-N2.5-mini mlx-community/Nex-N2.5-mini-OptiQ-4bit, mlx-community/Nex-N2.5-mini-oQ4 MLX 4bit 24GB Nex-N2.5-mini Agent Model for Long-Horizon Tasks: 16GB+ VRAM
2026-09-08 openbmb/MiniCPM5-2B openbmb/MiniCPM5-2B-MLX, mlx-community/MiniCPM5-2B-8bit MLX, MLX 8bit 4GB MiniCPM5-2B On-Device Model Strong in Code and Math: 4GB+ VRAM

“Smallest VRAM tier" is the smallest tier in the requirements table of each article (for other builds, of those builds; for image, video and audio models, always the article’s own table, which counts every component). Leave headroom for context length.

Last updated 2026-09-26 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables under “Our Coverage and Data" are assembled by code from our article log.