Best Local LLM for a 24 GB Mac

Illustration of a laptop with three memory sticks beside it

The best local LLM for a 24 GB Mac is gpt-oss 20B, a 13.8 GB download for step-by-step reasoning, coding and tool use. For images and everyday work, Gemma 4 12B (7.7 GB on a Mac) fits with plenty of room. All of these run on the Mac itself, with nothing sent to a vendor, in free apps such as Ollama and LM Studio.

Key facts

  • Install first: gpt-oss 20B, 13.8 GB download.
  • Room for a model: about 18 GB of 24 GB, so a file under about 14 GB runs comfortably.
  • App: Ollama or LM Studio, both free. The commands below are for Ollama.
  • Cost: nothing. Check each licence before business use.

On this page

Models that fit a 24 GB Mac

ModelGood forDownloadFit on 24 GBLicenceCommand
gpt-oss 20BStep-by-step reasoning, coding, tools13.8 GBComfortableApache 2.0ollama run gpt-oss:20b
Gemma 4 12BChat, coding, reads images7.7 GBComfortableApache 2.0ollama run gemma4:12b
Qwen3 14BCoding, quality9.3 GBComfortableApache 2.0ollama run qwen3:14b
Phi-4 14BCoding, quality9.1 GBComfortableMITollama run phi4
Ministral 3 14BChat, reasoning, reads images9.1 GBComfortableApache 2.0ollama run ministral-3:14b
Gemma 4 E4BEveryday chat, reads images, calls tools9.5 GBComfortableApache 2.0ollama run gemma4
LFM2 24B-A2BPrivate agents, tool calling14.4 GBTightLFM Open License (own terms)ollama run lfm2:24b-a2b

Comfortable means the file fits with room for a long conversation. Tight means it loads, but close other apps and keep chats short. Sizes are what Ollama downloads on a Mac, read from its registry on 21 November 2026. Where Apple's MLX build of a model exists, Ollama now uses it on a Mac, so some sizes differ from the ones on Ollama's website. The popular 26B and 27B models, Gemma 4 26B and Qwen3.8 27B, are 16.2 to 18.2 GB on a Mac: too big for the roughly 18 GB a 24 GB Mac gives a model.

Which one for which job

  • Reasoning, planning and maths: gpt-oss 20B.
  • Everyday chat and images: Gemma 4 12B, or Gemma 4 E4B for speed.
  • Coding: Qwen3 14B or Phi-4 14B.
  • Private agents that call tools: LFM2 24B-A2B.

To check any other model against your Mac, use the local LLM memory calculator. For every Mac size on one page, see the best local LLM for Mac.

Tested on my Mac

I ran two models on my own MacBook Air: Apple M5 with 24 GB, macOS 27.0, Ollama 0.40.17, plugged in. Each speed is the middle of three runs after a warm-up, from Ollama's own timings. A token is roughly three-quarters of a word.

ModelDownload on my MacAnswer speedFirst loadMemory in use
Gemma 4 E4B9.5 GB (MLX build)87 tokens a second5.7 seconds9.4 GB
Llama 3.2 3B2.0 GB50 tokens a second1.3 seconds3.1 GB

Once loaded, both started answering a short question in under a tenth of a second. Gemma 4 E4B ran on Ollama's MLX engine, which is built for Apple chips; Llama 3.2 3B ran on its standard engine.

Turn off Low Power Mode first. With it on, Llama 3.2 3B ran at 17 tokens a second, about a third of its normal speed. Switch it off in System Settings, under Battery, before you run a local model.

Method: temperature 0, the same 256-token answer to a small-business question each time, and an 8,192-token context. Tested 21 November 2026.

How much room a 24 GB Mac gives a model

A Mac shares its memory between the processor and the graphics side, and by default macOS lets the graphics side use about 75% of it. On a 24 GB Mac that is about 18 GB. The conversation needs 2 to 4 GB on top of the model file, so a file under about 14 GB runs comfortably and one up to about 16 GB runs if you keep chats short.

Longer documents and longer chats use more memory, so leave more room if you paste in long reports. Speed depends mostly on the chip: a Pro or Max chip moves data faster than a base chip, so the same model answers faster.

Too big for 24 GB

Gemma 4 26B (16.2 GB), Qwen3.8 27B (18.2 GB): these need more room than a 24 GB Mac gives. See the best local LLM for a 32 GB Mac.

Which app to run it in

Ollama runs models in the background for other apps and agents to use; the commands above install and start each model. LM Studio gives you a chat window and a model search. Both are free. See Ollama vs LM Studio if you can't decide.

Frequently asked questions

What is the best local LLM for a 24 GB Mac?

The best local LLM for a 24 GB Mac is gpt-oss 20B, a 13.8 GB download for step-by-step reasoning, coding and tool use. For images and everyday work, Gemma 4 12B (7.7 GB on a Mac) fits with plenty of room.

How big a model can a 24 GB Mac run?

By default macOS lets a model use about 18 GB of a 24 GB Mac. Allow 2 to 4 GB for the conversation, so aim for a model file under about 14 GB, or up to 16 GB with short chats.

Can a 24 GB Mac run Gemma 4 26B?

Not comfortably. Gemma 4 26B is a 16.2 GB download, more than the roughly 16 GB that leaves room for a conversation. It needs a 32 GB Mac.

Are these models free?

Yes, to download and run, with nothing sent to a vendor. Check the licence before business use: Apache 2.0 and MIT allow commercial use, while the Llama, LFM and NVIDIA licences add their own conditions.

Sources

  • Ollama library and registry: model pages, and the download size for a Mac (read 21 November 2026). On my M5 MacBook Air, Ollama 0.40.17 chose the MLX build for Gemma 4 E4B.
  • ModelFit's open hardware dataset (CC BY 4.0), updated 26 October 2026. Good-for labels adapted; fit worked out with my own memory rule.
  • Local LLM memory calculator, for the 75% memory rule

Other Mac sizes: 8 GB · 16 GB · 32 GB · 48 GB · 64 GB · 128 GB.

New models arrive every week. See Local AI This Week for the latest. Spotted something out of date? Tell me.

Scroll to Top