April 2, 2026 launch · local ops runbook

Pick the Gemma 4 model your hardware can actually run.

Answer two questions about your device and priority. The selector maps them to a model size, VRAM target, and first runtime, then hands you the matching download and command path.

Get my model + VRAM plan or browse all download options

Free, independent guide. No signup. Updated October 2026.

Apache 2.0 140+ languages 128K / 256K context Text + image Audio on E2B / E4B / 12B Ollama / LM Studio
Ops default
26B A4B

The official Get started guide says a good place to start is Gemma 4 26B A4B for many tasks.

Fastest local
E4B

E4B is the easiest local recommendation when you want a capable first run on laptop-class hardware.

Top-end quality
31B

31B is the dense quality-first model for stronger coding, reasoning, and long-context work.

Model cards

Gemma 4 models by fit, not hype

These cards cover four of the five Gemma 4 sizes (the 12B size is in the memory table below). They are the real homepage. They tell visitors who each model is for, what memory range to expect, and which runtime path makes the most sense first.

Memory ranges below use the official Q4_0 to BF16 numbers from the Google AI for Developers model overview.

E2B

Edge-first and smallest footprint

Best for phones, browsers, Raspberry Pi style experiments, and the cases where getting Gemma 4 onto the device matters more than maximum depth.

2.9 GB to 11.4 GB 128K context Audio support
E4B

The safest first local choice

Best for laptop users who want the smoothest first run in LM Studio or Ollama without dropping all the way to the smallest model.

4.5 GB to 17.9 GB 128K context Audio support
26B A4B

The official default for many tasks

Best for consumer GPUs and builders who want a real step up in reasoning without going straight to the heaviest dense model.

14.4 GB to 57.7 GB 256K context MoE
31B

Flagship quality when hardware allows

Best for workstations, deeper coding workflows, document-heavy reasoning, and anyone explicitly optimizing for stronger output over lighter setup.

17.5 GB to 69.9 GB 256K context Dense
Download hub

Open the right Gemma 4 path the first time

Users searching “Gemma 4 download” usually do not want a long explanation. They want to know whether they should open weights, a GUI runtime, a CLI runtime, or a hosted sandbox.

Official sources first, then runtime shortcuts.

Official weights

Hugging Face

Best when you want the canonical weights, model cards, and the broadest handoff into the open-model tooling ecosystem.

Open Hugging Face
Official distribution

Kaggle

Best when you want the official Google distribution path and a direct bridge into notebooks, experiments, and model assets.

Open Kaggle
Fastest CLI

Ollama

Best when you want the shortest local command line path. This is the most common first stop for self-hosted users.

Open Ollama
Fastest GUI

LM Studio

Best when you want a desktop app, model browser, and the easiest non-terminal first run on a laptop or workstation.

Open LM Studio
Hosted test

Google AI Studio

Best when you want to test 26B A4B or 31B before investing in local setup or deciding how much model quality you really need.

Open AI Studio
Hosted build path

Gemini API

Best when your real goal is application development with a hosted endpoint, not local inference on your own machine.

Open Gemini API docs
Memory planner

Check the official memory budget before you choose a runtime

This table is the page’s most practical anchor because it answers the question that blocks real adoption: can your machine carry the model you want, at the precision you want, with enough headroom left for context and runtime overhead?

Google's estimates include loading overhead and depend on the inference tool. They are not Ollama download sizes. Context, KV cache and other applications still affect available memory.

Model BF16 SFP8 Q4_0 Use this as
E2B 11.4 GB 5.7 GB 2.9 GB Smallest path for phones, edge, and browser-oriented experiments.
E4B 17.9 GB 8.9 GB 4.5 GB Best laptop-class starting point when you want credible local performance fast.
12B 26.7 GB 13.4 GB 6.7 GB Mid-size dense model with text, image and audio input and a 256K context window; sits between the edge models and 26B A4B.
31B 69.9 GB 34.9 GB 17.5 GB Flagship dense model for quality-first reasoning and coding.
26B A4B 57.7 GB 28.8 GB 14.4 GB Officially recommended good place to start for many tasks when you can afford the extra memory.

Fast hardware rule of thumb

Compare the exact runtime package before committing memory. For example, Ollama's E4B Q4_K_M package is about 6.6 GB, while the E4B MLX package is about 9.5 GB. Neither is the total memory needed to run it. Keep space for your operating system, runtime and context; the theoretical Google Q4_0 figure is not an install-size promise.

What changes the real answer

Context length, KV cache, serving stack, and prompt size all move actual usage. If you are near the limit, test the runtime early instead of trusting a base-weight number as a production guarantee.

Quick commands

Copy the shortest Gemma 4 starting points

This section is intentionally biased toward action. If a visitor already knows the model they want, the fastest way to keep them on the page is to let them copy something useful immediately.

Check the current runtime catalog before production use. Launch support changes quickly.

Ollama

Run the selected local model

The command follows your selector choice. Install and open Ollama first.

ollama --version
ollama run gemma4:e4b-it-q4_K_M "Reply with one short greeting."
ollama ps

Laptop example, reviewed November 2, 2026: Apple Silicon Mac, macOS 14 or newer, 16 GB or more unified memory, and a short 4K-context text test. This is a starting configuration, not a measured speed or minimum-memory guarantee. The selected E4B package uses Q4_K_M weights with a BF16 projector and is about 6.6 GB. Ollama v0.40.13 was the current stable release checked. Use ollama ps to inspect memory, processor and context; reduce context or try a smaller model if memory is tight.

Check the exact package, reviewed runtime release, Mac requirements and context settings. This text route does not verify audio input or long-context performance.

Python API

Test hosted access with Python

Best when you want to try Gemma 4 quickly through the Gemini API before self-hosting.

from google import genai

client = genai.Client(api_key="YOUR_API_KEY")
response = client.models.generate_content(
    model="gemma-4-31b-it",
    contents="Roses are red..."
)
print(response.text)
REST API

Call Gemma 4 over REST

Best when you want the raw hosted request path without installing an SDK first.

curl "https://generativelanguage.googleapis.com/v1beta/models/gemma-4-31b-it:generateContent?key=YOUR_API_KEY" \
  -H 'Content-Type: application/json' \
  -X POST \
  -d '{
    "contents": [{
      "parts": [{"text": "Roses are red..."}]
    }]
  }'

Next after quick start

Once Gemma 4 is running, your next question is usually not “what is Gemma 4?” anymore. It becomes “should I move up a model size, stay local, or shift to a hosted API?” That is why the two most important follow-up pages are model sizes and pricing/license.