Blog

How to run DeepSeek locally on a Mac with MLX — open reasoning (R1) and coding models from a 4 GB 8B on a MacBook Air to the V4-Flash flagship on a Studio, one command to serve them, and how to wire them into Claude Code or Cursor.

How to run Google's Gemma locally on a Mac with MLX — which size fits your RAM (from a 1 GB nano on an 8 GB Air to the 31B dense on a Studio), one command to serve it, and how to wire it into Claude Code or Cursor.

How to run OpenAI's GPT-OSS locally on a Mac with MLX — the 20B fits a 16 GB MacBook, the 120B runs on a Studio, both mixture-of-experts with native reasoning and tool calling. One command to serve it, and how to wire it into Claude Code or Cursor.

How to run Meta's Llama locally on a Mac with MLX — which size fits your RAM (from a 0.5 GB 1B on a base Air to the 8B on 16 GB), one command to serve it, and how to wire it into Claude Code or Cursor.

How to run Mistral locally on a Mac with MLX — Mistral Small 24B on a 24 GB MacBook, Devstral for coding, and the Mistral Small 4 119B flagship on a Studio. One command to serve it, and how to wire it into Claude Code or Cursor.

MLX vs llama.cpp performance on Apple Silicon, measured on the same machine with the same models — llama-bench vs mlx-lm plus full server-path numbers. MLX wins decode (1.5× on MoE), llama.cpp wins prefill (1.7–2×), and which one is "faster" depends on your workload shape.

We benchmarked rapid-mlx against Ollama, llama.cpp and LM Studio on an M2 Pro Mac mini — three models, same weights where possible, medians of three runs, raw JSON published. Rapid-MLX wins MoE decode by 1.5×, loses cold prefill, and one engine surprised us.

The best local AI models for Apple Silicon in 2026, picked by job — best all-rounder, best for coding agents, best under 8 GB, best reasoning, best vision, best flagship — with real sizes, real RAM requirements, and one command to run each.

How to run Qwen 3.5 and Qwen 3.6 locally on a Mac with MLX — which size fits your RAM (from 8 GB Airs to 128 GB Studios), one command to serve it, and how to wire it into Claude Code or Cursor.

One pick per memory tier, from an 8 GB MacBook Air to a 96 GB Studio. Boot time and peak memory measured in a single run, capability scored against the full-precision base models, and the one place our own table does not climb.

82 models plotted against the unified memory they actually need on Apple Silicon, with Artificial Analysis intelligence scores where they exist. 64 of them fit in 24 GB. The interesting part is what does not fit anywhere.

We re-verified Claude Code, Codex CLI, Aider, and Hermes end-to-end on local models on Apple Silicon (July 2026) — real edits, real tool calls, current binaries. What works, why, and the one-line setup for each.

An honest, reproducible comparison. Where each one fits, real rapid-mlx tok/s on an M3 Ultra, and a copy-paste benchmark to measure all three on your own Mac — because we don't publish head-to-head speed numbers we can't reproduce.

Fifteen releases in three weeks. Prefix-cache reuse takes TTFT from 13.1s to 0.51s, rapid-mlx chat becomes a real MCP agent, five new model families land day one, and tool calls are now schema-valid by construction — all measured on real hardware.

A practical migration guide. The CLI commands map almost 1:1 (run, pull, rm, ps), you point your OpenAI-compatible tools at a new port, and you can run both servers side by side while you switch. Includes a model-name cheat sheet.

Point Aider at a model running locally with rapid-mlx. Two env vars, and the terminal pair-programmer edits your files entirely on your Mac — private, offline, $0 per token. Works even with reasoning-only models.

Point Anthropic's Claude Code at a model running on your own Mac with rapid-mlx. Two environment variables, 100% local, $0 per token, and tool calls that actually work — verified on rapid-mlx 0.11.1. Here's the whole setup.

Point OpenAI's Codex CLI at a model running locally with rapid-mlx. One config block (or a one-shot --setup), and Codex talks to your Mac the same way it talks to api.openai.com — private, offline, $0 per token.

Use Cursor with a local LLM on your Mac — point chat, Ctrl-K, and composer at a model served by rapid-mlx. One custom base URL in Settings → Models, and your editor's AI runs on your Mac: private, offline, $0 per token.

I benchmarked five local LLMs that fit a MacBook Pro — Gemma 4, Qwen3.6, GPT-OSS-20B, Nemotron-Nano-30B, and a ternary Bonsai-27B — on coding, reasoning, tool-calling, and memory. One of them is a 27B model squeezed into 8GB, and it nearly won.

Standardized HumanEval+, MATH-500, MMLU-Pro and tool-calling evals across 17 MLX models on a 256 GB M3 Ultra — decode speed, RAM, and accuracy for every model, plus what the community found, so you can pick the best local LLM for agent and coding work on Apple Silicon.