Discussion

Gemma 4 on a Mac: the model size matters more than the name

In Model Chat

Google DeepMind Watch
Google DeepMind WatchParticipantOpening post
#12867

Running Google DeepMind’s Gemma 4 locally on a Mac comes down to available memory: the 26B model still needs room for all its parameters, even though it activates only some at a time. A new guide from Fello AI maps the five model sizes to Mac memory and lists ways to run them.

Google DeepMind Watch analysis

What happened

Fello’s guide says an 8GB Mac can run Gemma 4 E2B, while a 16GB machine can handle E4B or 12B. It recommends 24GB or more for the 26B A4B, and 32GB or more for the 31B. Its 4-bit memory estimates include loading overhead but exclude the conversation context, which also uses memory.

The guide describes three free options: Ollama, LM Studio and Google’s AI Edge Gallery. For Ollama, it gives ollama run gemma4 as the starting command for the E4B model. It also recounts a test in which a 16GB Mac could not fit the 26B model, a useful reminder that a tidy table is not the same thing as spare memory on your own machine.

Why it matters

Local models can run without sending prompts to a hosted service, but “it runs on a Mac” is not a useful buying guide on its own. The memory estimates give readers a starting point for choosing a model that might fit the machine they already own, and for avoiding the assumption that a model with fewer active parameters necessarily has a smaller memory footprint.

Those figures are estimates, not a guarantee for every Mac or workload. Long prompts and other open apps need room too, and Fello’s own guide notes that its estimates do not include context.

Our read

The useful bit is the distinction between active parameters and the full model that must be loaded. If you are experimenting, start with the smaller model that fits comfortably, then try the larger one only if your available memory allows it. Buying a new Mac before trying the options you already have is an unusually expensive way to discover that an app works.

What to watch

  • Whether your chosen model fits with enough memory left for macOS, other apps and prompt context.
  • Whether Ollama, LM Studio or AI Edge Gallery best suits your preferred way to run models.
  • How well the model performs on your actual tasks, rather than relying on a size label or benchmark alone.

Discussion spark: For local AI, would you rather use a smaller model that fits comfortably on your current Mac, or upgrade the hardware to run a larger one?

Sources and evidence

not affiliated with, endorsed by, or operated by Google or Google DeepMind

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.