Gemma 4: Google’s Free Open Model, and Which Mac Can Run It

Gemma 4 is Google DeepMind’s family of free, open-weight AI models, announced on 2 April 2026 under the Apache 2.0 licence. It now comes in five sizes, E2B, E4B, 12B, 26B A4B and 31B, and every one of them reads text and images and writes text. You download the weights once and run them on your own phone, laptop or server, with no subscription and no internet connection.

The hard part is not getting Gemma 4. The hard part is picking the size your machine can actually hold.

The names hide the memory bill. The “4B active” 26B model still needs room for all 26 billion parameters. This guide maps each Gemma 4 model to the Mac that runs it, using Google’s own memory figures. It then covers three ways to install it, its scores, how it stacks up against Qwen 3.6, and what it cannot do.

The Key Takeaways

• Five sizes, all free: E2B, E4B, 12B, 26B A4B and 31B, under Apache 2.0, which allows commercial use.
• Memory decides the model: at 4-bit, Google lists 2.9GB for E2B, 6.7GB for the 12B and 17.5GB for the 31B, before any context.
• The Mac short version: 8GB runs E2B, 16GB runs E4B or the 12B, the 26B A4B wants 24GB or more, and the 31B is happiest at 32GB+.
• Strong for its size, not a frontier model: the 31B scores 84.3% on GPQA Diamond, but Alibaba’s own table puts Qwen3.6-27B ahead on coding, 77.2% vs 52.0% on SWE-bench Verified.
• Reads, does not draw: Gemma 4 takes images (and audio on the small models) as input but only writes text, and its training data stops in January 2025.

What Is Gemma 4?

Gemma 4 is Google DeepMind’s open model family, the downloadable sibling of Gemini. Google’s launch post says it is “built from the same world-class research and technology as Gemini 3.” The same post calls it “the most capable model family you can run on your hardware.” Gemini stays a closed service you reach through Google’s apps and API. Gemma 4 ships as weights you keep.

The family has grown in three steps. Google’s release log dates the first four models, E2B, E4B, 26B A4B and 31B, to 31 March 2026, and the public announcement followed on 2 April. Multi-Token Prediction (MTP) drafters for those four arrived on 16 April. Gemma 4 12B Unified, a laptop-sized middle model with native audio input, completed the line-up on 3 June 2026.

Gemma 4 landed well. At launch, Google said the 31B ranked as the #3 open model on the Arena text leaderboard and the 26B as #6. By 3 June, Google counted more than 150 million downloads of Gemma 4 models. The licence matters as much as the scores. Earlier Gemma generations shipped under Google’s own Gemma Terms of Use. Gemma 4 uses plain Apache 2.0, a licence much of the open-source world already understands. For how that compares with the custom licences on other open models, see our ranking of the best open source AI models.

Gemma 4 Models: All Five Sizes Compared

Gemma 4 comes in five sizes, and the smallest and largest differ by more than ten times in parameters. Parameters, context and inputs come from Google’s model card. The 4-bit (Q4_0) figure comes from the Gemma 4 memory table in Google’s documentation, which counts the weights plus 20% loading overhead but not the conversation itself.

What E2B and E4B Mean

The “E” in Gemma 4 E2B and E4B stands for “effective” parameters. Google built both for phones and edge devices using Per-Layer Embeddings, which give every layer its own small lookup table instead of adding more layers. The lookup tables are large, so E4B behaves like a 4.5B model in speed but loads 8B parameters in total. Plan memory for the bigger number.

Why the 26B A4B Needs 26B Worth of Memory

Gemma 4 26B A4B is a Mixture-of-Experts model. It holds 128 experts and switches on 8 of them, plus one shared expert, for each token. Only 3.8 billion parameters do the work at any moment. That makes it fast, close to a 4B model in speed. It does not make it small. Google’s documentation says “all 26 billion parameters” must be loaded into memory to keep the routing fast, so its memory need sits close to a dense 26B model. The 4-bit figure is 14.4GB, against 4.5GB for E4B.

What 12B Unified and MTP Changed

Gemma 4 12B Unified, released on 3 June 2026, drops the separate vision and audio encoders the other models use. Images and sound go straight into the language model through light projection layers. Google says the 12B delivers benchmark performance “nearing” the 26B model at less than half its memory, and that it runs on laptops with 16GB of RAM. It is also the first mid-sized Gemma 4 model to take audio.

Multi-Token Prediction is the other addition, and it is about speed.

Every Gemma 4 model now ships with a small “drafter” that guesses several tokens ahead while the main model checks them in one pass. Google says the output is exactly the same as normal generation, only faster. The gain is smallest on the 26B A4B, because checking drafted tokens there can pull extra experts into memory.

Which Gemma 4 Model Can Your Mac Run?

Any Apple Silicon Mac can run a Gemma 4 model, and unified memory decides which one. The table below starts from Google’s 4-bit and 8-bit figures and Ollama’s actual download sizes, which run a little larger. Leave several gigabytes free for macOS and your open apps, and more again for long prompts: Google’s figures exclude the context, which grows with every token in the conversation.

Real tests line up with the table. A developer at Classmethod tried the 26B on a 16GB M5 MacBook, where Ollama reported it needed 18.8GiB against about 11.8GB available to the model, and settled on E4B. If you are choosing hardware rather than a model, our guide to the best MacBook for AI lists which configurations reach 24GB and 32GB. Our M5 Ultra Mac Studio review shows the Gemma 4 26B A4B loading in 15.6GB with room to spare.

When a Local Model Is Not Enough

Gemma 4 on a 16GB Mac is a good private, offline assistant for drafts, summaries and reading screenshots. A 12B model is not what you want for a hard coding problem or a long legal document. You do not have to buy a bigger Mac to fix that. With Fello AI, you keep Gemma for the offline work and ask Claude, GPT or Gemini the hard question in one app, without a second subscription or more RAM.

Google’s own Gemma team makes the same split. Its account described the efficient setup this way on 2 November 2026:

The most efficient startup architectures run a compound AI stack: pairing frontier APIs for heavy reasoning with open weights for fast, high-volume tasks.

Perfect fits for your stack:

📱 Edge & local execution
🚦 Agent routing & triage
🛠️ Custom fine-tuning
🏥 Vertical domains (like MedGemma)

— Google Gemma (@googlegemma) November 2, 2026

How to Run Gemma 4 on a Mac

The fastest way to run Gemma 4 on a Mac is Ollama, which takes one command after install. LM Studio is the choice if you prefer a normal app window, and Google’s own AI Edge Gallery is the most hands-off option. All three are free.

Option 1: Ollama

1. Install Ollama. Its macOS download page gives a one-line Terminal command: curl -fsSL https://ollama.com/install.sh | sh.
2. Run the default model. In Ollama, the plain gemma4 tag is E4B, the right start for a 16GB Mac:

ollama run gemma4

3. Pick a different size by adding its tag. The Ollama Gemma 4 library page lists five: gemma4:e2b, gemma4:e4b, gemma4:12b, gemma4:26b and gemma4:31b.

ollama run gemma4:12b

4. Chat in the Terminal, or point any app that speaks the OpenAI API format at Ollama’s local server, http://localhost:11434/v1/.

Ollama also lists MLX builds for Apple Silicon, such as gemma4:e4b-mlx, and a gemma4:31b-cloud tag that runs the 31B on Ollama’s servers instead of your Mac. Beware of guides that tell you to pull gemma4:4b or an “E12B”: neither tag exists.

Option 2: LM Studio

LM Studio runs Gemma 4 in a chat window with no Terminal. Search for Gemma 4 in its model browser, choose the GGUF or the MLX build (MLX is the Apple-native format), download, and start chatting. LM Studio says the smallest Gemma 4 needs at least 4GB of RAM and the largest up to 19GB.

Option 3: Google AI Edge Gallery

Google AI Edge Gallery is Google’s own app for running Gemma on-device. Google brought it from iPhone to the Mac alongside Gemma 4 12B in June 2026, as a free direct download from its developer site. It runs the 12B offline on the Mac’s Apple Silicon GPU. It also includes a sandboxed Python loop, so Gemma 4 can write and run code to plot a chart inside the chat.

Is Gemma 4 Free?

Yes, Gemma 4 is free in both of the ways that matter. The weights are free to download from Hugging Face, Kaggle, Ollama and LM Studio, and Apache 2.0 lets you use them commercially, change them and ship them in a product. Running them locally costs only electricity.

Google also hosts two Gemma 4 models for free through the Gemini API: gemma-4-31b-it and gemma-4-26b-a4b-it. Google’s pricing page lists input, output and context caching as “Free of charge” and shows no paid tier at all. The catch is data: on the free tier, Google marks your prompts as “Used to improve our products”. Our guide to free AI models compares that route with the other no-cost options. If privacy is the reason you want Gemma, run it on your own Mac instead.

Gemma 4 Benchmarks: How Good Is It?

Gemma 4 31B is strong for a model that fits on a laptop, and the jump over Gemma 3 is large. The scores below come from Google’s model card and cover the instruction-tuned models with reasoning on.

Two readings matter. First, the 26B A4B stays within three points of the 31B on most rows while running at roughly 4B speed, which makes it the sweet spot for a 32GB Mac. Second, the edge models fall off sharply on maths and coding: E4B is fine for summaries and reading text in images, not for competition maths. Independent scores are thin. As of 4 November 2026, Artificial Analysis lists Gemma 4 models but publishes no Intelligence Index v5 score for any of them, so treat the vendor numbers as the ceiling.

Gemma 4 vs Qwen: Which Open Model Should You Pick?

Qwen 3.6 beats Gemma 4 on coding and edges it on reasoning, according to Alibaba’s own benchmark table. That table is useful because it runs Gemma 4 31B in the same columns, and its Gemma figures match Google’s model card exactly. The numbers are still vendor-reported by Alibaba, so read the margins, not the decimals.

Source: the Qwen3.6-27B model card on Hugging Face. The SWE-bench Verified gap, 77.2 against 52.0, is the one that changes a decision.

If you want a local model to edit real code across a repository, Qwen3.6-27B is the stronger pick. Gemma 4 keeps three advantages. Apache 2.0 covers the whole family. Its small edge models take audio, a job Qwen’s 27B does not cover. And Google ships first-party apps for it on the Mac and iPhone. Our list of the best Chinese AI models covers the newer Qwen 3.8 and its rivals.

What Gemma 4 Cannot Do

Gemma 4 cannot generate images. Every model in the family takes images in and writes text out. It can describe a photo, read a chart or pull text from a scanned page, but it cannot draw one.

Its knowledge is old. Gemma 4’s training data has a January 2025 cutoff, so it does not know recent events or newer models unless you paste the facts in.

Audio and video are limited. Only E2B, E4B and 12B take audio, capped at 30 seconds per clip, and the 26B A4B and 31B do not hear at all. Ollama’s library lists text and image input only. Video is processed as frames, up to about 60 seconds at one frame per second.

Context costs memory. The 128K and 256K windows are real, but filling them adds memory on top of the figures in the tables. That is where 16GB Macs run out first.

It is not a frontier model. The 31B scores 19.5% on Humanity’s Last Exam without tools in Google’s own card, far below the closed flagships.

The Bottom Line

Gemma 4 is the best reason yet to run an AI model on your own Mac. Pick by memory, not by name: E4B or the 12B on 16GB, the 26B A4B on 32GB, and the 31B only when you have room to spare. Choose Gemma 4 for private, offline work, image reading and a licence your company’s lawyers already know, and choose Qwen 3.6 if local coding is the main job. Start with ollama run gemma4, see how your Mac copes, and size up from there.

FAQ