How to run Gemma 4 on a Mac: RAM, speed and where it falls short

I ship Gemma 4 inside Outlier as the Quick tier, so I have timed it and tested it more than most people. This page is what I measured on one Mac Studio, which Macs have room for it, three ways to run it, and the two places it let me down. If you only want the commands, jump to "Three ways to run it".

Which Mac has room for Gemma 4

Gemma 4 comes in five sizes: E2B, E4B, 12B, 26B-A4B and 31B. The Gemma vs Llama page has the launch details and the licence. I measured one of them, the 26B-A4B. It is a mixture-of-experts model. The file holds 26 billion parameters, but only about 4 billion of them are used for each word it writes. At 4-bit in MLX format the file is 15.61 GB. While it answers, it peaks near 16 GB of memory.

Table 1: Gemma 4 26B-A4B at 4-bit, by Mac memory

The middle column is just the memory minus 16 GB. I did not time a 16 GB Mac for this page. For the general rule, see how much RAM local AI needs.

If you have 8 GB or 16 GB, the free tiers fit better. Outlier's Nano is a 2.37 GB file and Lite is 5.04 GB. They are Qwen models, not Gemma. If you want Gemma itself on a small Mac, E2B and E4B are the sizes Google makes for that. I have not measured them, so I can't give you a speed.

Three ways to run it

1. With MLX, free, from the Terminal

This is the same library Outlier uses to load Quick. It needs a Python environment.

pip install -U mlx-lm
mlx_lm.generate --model mlx-community/gemma-4-26b-a4b-it-4bit \
  --prompt "Explain unified memory in two sentences."

The first run downloads about 15.6 GB from Hugging Face. After that it loads from disk. For a back-and-forth chat, use mlx_lm.chat with the same --model. Outlier ships mlx-lm 0.31. If your copy is older and won't load the model, update it first. The Hugging Face repo is tagged Apache 2.0.

2. In Outlier, as the Quick tier

Quick is Gemma 4 26B-A4B as a 4-bit MLX file. It is part of Pro. The free tiers are Nano and Lite.

1. Download and open Outlier.
2. Open the model menu at the top of the chat and pick Quick. If it isn't on your Mac yet, you download it from that same menu (15.61 GB).
3. If your Mac has less memory than Quick's 16 GB minimum, the menu warns you before it downloads anything.

Once it is downloaded it answers with the Wi-Fi off.

3. In Ollama or LM Studio

Google's own docs list both as ways to run Gemma 4. I have not tested Gemma 4 in either, so I won't give you a command that I haven't run. The steps are the same as for any other model: search for Gemma 4 in LM Studio's model browser, or pull it by name in Ollama. How to run Llama on a Mac walks through both apps, and Ollama vs LM Studio helps you pick.

How fast it is

Table 2: decode speed on an M1 Ultra, 4-bit MLX, one request at a time

The rows were measured on different days and builds, so read the ratios, not the third digit. The full reference table, with the method, is on Mac RAM to AI model size.

Why is Quick about twice as fast as Core when the files are the same size? On a Mac, writing each word is mostly limited by how quickly the chip can read the model's weights from memory. Quick reads about 4 billion parameters' worth per word. Core reads all 27 billion. Quick is still a little slower than the 9B Lite, which I put down to the routing work a mixture-of-experts model does on every word. I haven't isolated that.

One more thing about speed. Quick reasons before it answers. In my knowledge test it wrote about 146 reasoning tokens before each answer letter. At 43.6 tokens a second that is roughly 3 seconds of thinking before the first line shows up. That is arithmetic from two measured numbers, not a timing of its own. I did not time first-word latency on Quick.

How good it is

Table 3: Quick (Gemma 4 26B-A4B), my own tests

Knowledge. On the MMLU sample, Quick and the 9B Lite are tied inside the noise. If all you need is facts and explanations, the free 5 GB Lite is a fair alternative. Quick's score came from reading an answer letter out of a chat reply, which is noisier than the standard scoring and can miss right answers. Lite's number is from the same sprint's table, and the two may not be scored the same way.

Short code. Asked for a single function with the problem already stated, Quick matched the 27B Core and finished in 37% less time. Between them they solved all 23, because their misses did not overlap. The sample is small, so one point either way is noise.

Real repositories. This is where it falls down. Given a bug in a real project and left to find the fix itself, Quick produced no usable patch in 29 of 30 tries. It stalled while searching and never settled on an edit. I wrote that test up in chat coding vs agentic coding. Core stays the default for agent work for that reason.

HumanEval. An older run in April scored Quick 21 of 164, with a harness that pulls the code out of a fenced block. That disagrees with the 21 of 23 above, and I haven't worked out why. Treat Quick's coding as short-function help, and use Core for anything larger.

When I would pick it

On a 24 GB or larger Mac, for quick answers, drafting and explaining things, Quick is a good fit, and it is the faster of the two big models. For code inside a project, use Core. On a 16 GB Mac, start with Lite, which is free. If you aren't sure local AI is for you, download the free app and try Nano and Lite first. They cost nothing.

What I did not test

• A 16 GB Mac, or any chip other than an M1 Ultra. The speed on an M4 will differ.
• The other four Gemma 4 sizes (E2B, E4B, 12B and 31B).
• Gemma 4 in Ollama, LM Studio or llama.cpp.
• Images, audio or video. I only tested text.
• The two Terminal commands above. Outlier loads Quick through the same mlx-lm library, but I did not retype them for this page.
• First-word latency on Quick.

Frequently asked questions