Blog

Best Local LLMs for a MacBook Pro (2026)

Five models that fit a MacBook Pro, put through the same coding, reasoning and tool-calling probes. Then the part that decides whether you'll enjoy one: how fast each actually runs on Pro and Max chips, measured on those chips.

· updated Oct 12, 2026 · 9 min read
benchmarksmlxapple-siliconmacbook-pro

Your MacBook Pro is a capable private AI machine: no account, no monthly bill, nothing you type leaving the laptop. The real question is which model to run, and most advice out there is either vibes or numbers from a Mac Studio that looks nothing like a laptop.

So this post does two things. It scores five local models that fit a MacBook Pro on the same coding, reasoning, general-knowledge and tool-calling probes. And it shows how fast they run on the chips MacBook Pros actually ship with, using speeds measured on those chips by the Rapid-MLX community and by us, never scaled from a bigger Mac.

Quick answer: which model for a MacBook Pro?

These are the engine's v0.16.0 general-purpose picks by unified memory. A 36 GB MacBook Pro uses the 32 GB tier; 128 GB uses the 96 GB tier. The five-model comparison below is a separate set of measured alternatives.

Mac memory Recommended alias Command Good for
8 GB lfm2.5-2.6b-4bit rapid-mlx serve lfm2.5-2.6b-4bit Chat and tools; avoid coding
16 GB qwen3.5-4b-4bit rapid-mlx serve qwen3.5-4b-4bit Everyday chat and code
24 GB bonsai-27b-2bit rapid-mlx serve bonsai-27b-2bit Larger text-only tasks
32/36 GB qwen3.8-27b-4bit rapid-mlx serve qwen3.8-27b-4bit General work and code
48 GB qwen3.8-27b-4bit rapid-mlx serve qwen3.8-27b-4bit General work with headroom
64 GB qwen3.8-27b-4bit rapid-mlx serve qwen3.8-27b-4bit General work with headroom
96/128+ GB qwen3.8-27b-4bit rapid-mlx serve qwen3.8-27b-4bit General work with more headroom

At 48 GB and above, the engine also offers rapid-mlx serve qwen3.6-35b-4bit as its fast pick. For a broader memory guide, see best local LLM by RAM, how much memory a local LLM needs, and the model catalog.

How the five scored

Each model ran the same probes with thinking off and temperature 0. The percentage is the average of four suites, followed by each model's strongest.

Quantization: Gemma, Qwen and Nemotron at 4-bit, GPT-OSS at MXFP4, Bonsai ternary (under two bits per weight).

One honest warning. Each capability score is a 10-question probe (30 scenarios for tools), not a full academic benchmark. That is enough to sort these five into tiers and catch a model that face-plants. It is not enough to settle a three-point gap. Treat ±10% as noise and the tiers as real.

Speed on real MacBook Pro chips

These chips are the ones MacBook Pros ship with; a few of the runs came from a Mac mini or Studio with the same chip and memory. Decode speed is how fast the model writes once it has read your prompt. Each row is one measurement, with the rapid-mlx version it ran on. "Our run" rows are from our own Mac mini on the model's default settings, "older" rows from the previous community benchmark, and the rest from the current one.

Gemma-4-26B

Run it with rapid-mlx serve gemma-4-26b-4bit.

Peak memory: 16.9 GB in the community run, 20.8 GB Metal peak in ours. Plan on a 32 GB Mac.

Qwen3.6-27B

Run it with rapid-mlx serve qwen3.6-27b.

It's the only dense model of the five, so every token touches all 27 billion weights. That's why it's the slowest here, and why a Max chip's extra memory bandwidth helps it most. Peak memory: 18.1 GB in the M4 Pro community run, 22.8 GB Metal peak in ours.

GPT-OSS-20B

Run it with rapid-mlx serve gpt-oss-20b.

Peak memory is 12.4 GB in the community runs, which makes it the lightest "genuinely useful" model here.

Bonsai-27B

Run it with rapid-mlx serve bonsai-27b-2bit.

A 27-billion-parameter model squeezed to under two bits per weight, and it still scored 81%. On a laptop it's slower than its size suggests, and it uses more memory than its 8 GB of weights: 11.0 GB peak in the community run, 13.5 GB Metal peak in ours. That's why the engine places it in its 24 GB tier, not at 16 GB.

Nemotron-Nano-30B

No runs on a Pro or Max chip yet. It also scored lowest, partly because these probes ran with thinking off, which hurts a think-first model most.

Every number above comes from the leaderboard data or our own runs, and each appears on a page per Mac: see M4 Pro with 48 GB, M2 Pro with 32 GB, M1 Max with 64 GB and every Mac with enough runs.

Which one fits your MacBook Pro?

rapid-mlx recipe reads your Mac's memory and prints the engine's pick with the command to run it. The picks below are that policy, the same one the installer and the desktop app use. After each, which of the five fit, going by the peak memory measured above.

The engine's general-purpose pick, Qwen3.8-27B, came out after these five were scored, so it isn't in the capability table. It's measured on Pro and Max chips though: 32.9 tok/s on an M5 Max with 64 GB, 18.0 on an M5 Pro with 64 GB, and 12.1 on an M1 Max with 64 GB.

Run any of these

Everything above runs on Rapid-MLX, a local server with OpenAI- and Anthropic-compatible APIs. Claude Code, Aider and other OpenAI-compatible tools point at it by changing one URL:

curl -fsSL https://rapidmlx.com/install.sh | bash
rapid-mlx recipe
rapid-mlx serve gemma-4-26b-4bit

Homebrew works too: brew install rapid-mlx. recipe prints the pick for your Mac; serve starts the OpenAI API at http://localhost:8000/v1 and the Anthropic API at http://localhost:8000.

If a model won't fit, serve says so before loading instead of thrashing your Mac. On the same model and weights, Rapid-MLX is up to 4× faster than mlx-lm.

The fine print

MacBook Pro model questions

What is the best local AI model for a 16 GB MacBook Pro?

The engine picks qwen3.5-4b-4bit for 16 GB. Run rapid-mlx serve qwen3.5-4b-4bit; see the RAM guide for other memory tiers.

What local LLM should I run on an M1 MacBook Pro with 32 GB?

The 32 GB recommendation is qwen3.8-27b-4bit regardless of chip generation. Run rapid-mlx serve qwen3.8-27b-4bit, and check memory requirements if you also keep other large apps open.

Which model should I use on a 48 GB MacBook Pro?

The general-purpose pick remains qwen3.8-27b-4bit: run rapid-mlx serve qwen3.8-27b-4bit. For the engine's fast alternative, run rapid-mlx serve qwen3.6-35b-4bit.

Can I choose a different MacBook Pro model?

Yes. Run rapid-mlx recipe for the default picks, then browse all supported models for a specific task; the measured models above are alternatives to rapid-mlx serve qwen3.8-27b-4bit on a 32 GB or larger Mac.

Run this yourself. Rapid-MLX is an open-source, OpenAI- and Anthropic-compatible inference server for Apple Silicon. One command installs it, then rapid-mlx serve <alias> serves any model on localhost:8000/v1.
Download for Mac Pick a model by your Mac's RAM 
curl -fsSL https://rapidmlx.com/install.sh | bash

New models and speedups, in your inbox

A short note whenever Rapid-MLX gets faster or adds models worth running on your Mac. No spam — unsubscribe anytime.