★ Reading this for free? Get 20 structured AI courses + per-chapter AI tutor — the first chapter of every course free, no card.Start free in 30 seconds

Google · Open Weights · Apache 2.0

Gemma 4: Every Model, and What Hardware Actually Runs It

Gemma 4 (launched March 31 / April 2, 2026, Apache 2.0) is the most consumer-runnable open model family of the summer 2026 releases. Five members, five hardware tiers: E2B and E4B run fully offline on phones and Raspberry Pi (Google's claim); the 12B "Unified" — added June 3 — takes text, audio, and image in one model with 256K context and fits a 12GB card at Q4 (~7-8 GB, our estimate); the 26B-A4B MoE and 31B dense fit a single 24GB card quantized (our estimate), and Google says both fit a single 80GB H100. Below: every family member mapped to real hardware, with every estimate labeled as one.

📅 Published: July 20, 2026🔄 Last Updated: July 20, 2026✓ Manually Reviewed

Key takeaways

  • Five models, phone to H100 — E2B (~2.3B effective), E4B (~4.5B effective), 12B Unified, 26B-A4B MoE, 31B dense.
  • Apache 2.0, open weights — no custom license restrictions.
  • 12B Unified is the headline — text + audio + image input in one model, up to 256K context. Added June 3, 2026.
  • Phone and Pi are official targets — Google claims E2B/E4B run fully offline on phones, Raspberry Pi, and Jetson Orin Nano.
  • 24GB VRAM unlocks the whole family — 26B-A4B at Q4 ≈ ~15-16 GB, 31B at Q4 ≈ ~18-19 GB (our estimates).

Quick verdict

If you run models locally, Gemma 4 is the family to know in 2026. Many 2026 open releases target datacenter hardware; Gemma 4 goes the other way: every member of the family targets hardware people actually own, from phones up to a single prosumer GPU.

The short version: own a 12GB card, run the 12B Unified. Own a 24GB card, run 26B-A4B (faster) or 31B (denser). Own a Raspberry Pi or want offline-on-phone, E2B/E4B are Google's explicit targets. For install steps, our Gemma local setup guide covers the tooling.

The family at a glance

The core family launched March 31 / April 2, 2026 (Google's announcement). The 12B Unified joined on June 3, 2026, confirmed on Google's official Gemma releases page. All under Apache 2.0. Google's release notes document input modalities only for the 12B Unified (text + audio + image), so that's the only member where we list them.

ModelParametersNotable
E2B~2.3B effectivePhones, Raspberry Pi, Jetson Orin Nano — fully offline (Google's claim)
E4B~4.5B effectiveSame targets as E2B, more capable
12B Unified12BText + audio + image input in one unified model, up to 256K context. Added June 3, 2026
26B-A4B26B total / ~4B active (MoE)MoE speed: quality of its size class, compute of a ~4B
31B31B denseThe full-strength dense flagship; fits a single 80GB H100 (Google's claim)

Hardware tiers: what runs where

Quantized sizes below are our estimates from standard Q4 arithmetic — Google publishes hardware targets, not GGUF file sizes. Treat them as planning numbers, and leave headroom for context.

Your hardwareBest Gemma 4 fitNotes
Phone / Raspberry Pi / Jetson Orin NanoE2B, E4BGoogle's stated targets — fully offline
8GB VRAME4B comfortably; 12B Unified at Q4 is tight12B Q4 ≈ ~7-8 GB (estimate) — short context only on 8GB
12GB VRAM12B Unified at Q4Comfortable fit with context headroom (estimate)
16GB VRAM12B Unified with long-context headroom26B-A4B Q4 ≈ ~15-16 GB (estimate) — too tight once context is loaded
24GB VRAM26B-A4B or 31B at Q426B-A4B ≈ ~15-16 GB, 31B ≈ ~18-19 GB (estimates)
Apple Silicon, 32GB+ unified memory26B-A4B or 31B at Q4Unified memory covers the same estimates as a 24GB card
80GB H10026B-A4B or 31BSingle-GPU fit is Google's claim

On an 8GB card, E4B is the honest recommendation — the 12B squeeze works but leaves you no context room. More 8GB options in best local AI models for 8GB RAM. Sitting at 16GB and weighing alternatives? See the best LLMs for 16GB VRAM.

12B Unified: text, audio, image — one model

The June 3 addition is the most interesting member of the family. Gemma 4 12B "Unified" is a single unified multimodal architecture: it takes text, audio, and image input in one model, with up to 256K context. One model, one download — no separate vision or audio models to juggle.

For local use, that simplicity matters more than it sounds. Multimodal-local setups have historically meant juggling two or three models and hoping the glue code holds. A 12B that does all three inputs and fits a 12GB card at Q4 (~7-8 GB, our estimate) collapses that stack into one download. The 256K window is the other half of the story — enough for entire codebases or long transcripts, if your VRAM leaves room for it.

What defines the Gemma 4 lineup

Three structural choices define the Gemma 4 lineup. First, MoE: 26B-A4B carries 26B total parameters but activates only ~4B per token — its compute cost per token is closer to a small model's. Second, the E2B/E4B "effective parameter" models, built for phones and single-board computers as first-class targets rather than afterthoughts. Third, the unified multimodal 12B described above.

If you're currently running Gemma 2 9B, the 12B Unified is the natural upgrade path — similar hardware tier, plus audio and image input. If you're on Gemma 2 27B, the 26B-A4B gives you the same size class with MoE speed, and the 31B is the straight dense successor.

Ecosystem proof: local speech-to-speech in production

A fair question about any new family: is anyone actually building on it? For Gemma 4, yes — and the example is concrete. On July 1, 2026, Hugging Face and Cerebras open-sourced a complete local speech-to-speech pipeline: Nvidia Parakeet handles speech recognition, Gemma 4 31B is the language model in the middle, and Qwen3-TTS produces the spoken reply.

That pipeline isn't a demo repo — it already powers 9,000+ Reachy Mini robots. A fully local, open-weight voice assistant stack with Gemma 4 at its core, shipping in physical products three months after the family launched. That's the kind of ecosystem signal that usually takes a year.

FAQ

What is Gemma 4?
Gemma 4 is Google's open-weight model family, launched March 31 / April 2, 2026 under Apache 2.0. The lineup: E2B (~2.3B effective parameters) and E4B (~4.5B effective) for phones and single-board computers, a 26B-A4B Mixture-of-Experts model (26B total, ~4B active per token), and a 31B dense model. On June 3, 2026 Google added Gemma 4 12B "Unified" — a single model that accepts text, audio, and image input with up to 256K context. Everything ships as open weights you can download and run locally.
Can Gemma 4 run on a Raspberry Pi?
Google's claim is yes: E2B and E4B run fully offline on phones, Raspberry Pi, and Jetson Orin Nano. That's the vendor's own positioning, but the effective parameter counts (~2.3B and ~4.5B) make it plausible — these are the sizes that have historically been fine on Pi-class hardware. The bigger family members (12B Unified, 26B-A4B, 31B) are not Pi material; they need a real GPU or Apple Silicon with enough unified memory.
How much VRAM does Gemma 4 12B need?
Our estimate: roughly 7-8 GB at Q4 quantization. That fits a 12GB card comfortably with context headroom, and squeezes onto an 8GB card if you keep context short. Note the 12B Unified supports up to 256K context — long context costs additional memory on top of the weights, so a 12GB or 16GB card is the practical floor if you actually want to use that window. These are quant-size estimates, not official Google figures.
Should I upgrade from Gemma 2 to Gemma 4?
For most Gemma 2 users, yes. Three things stand out in the Gemma 4 lineup. First, a Mixture-of-Experts option: 26B-A4B runs 26B total parameters but only ~4B active per token, so it is fast for its quality class. Second, the E2B/E4B "effective parameter" models built explicitly for phones and Raspberry Pi (Google's claim). Third, the 12B Unified (added June 3, 2026): one model, one download, taking text, audio, and image input with 256K context — no separate vision or audio models to juggle. If you are on Gemma 2 9B or 27B today, the 12B Unified and 26B-A4B are the natural upgrades.
Is Gemma 4 open source? What license does it use?
Gemma 4 ships under Apache 2.0 — a genuinely permissive license, not a custom "open-ish" one. That covers commercial use, modification, and redistribution. The weights are open and downloadable. This is one of the reasons Gemma 4 matters for local AI: it is the most consumer-runnable open family of the summer 2026 releases, and there is no license gotcha attached.
Can Gemma 4 do speech / audio locally?
Two answers. The 12B Unified accepts audio as a native input alongside text and images. And for full voice conversations, there is working ecosystem proof: on July 1, 2026, Hugging Face and Cerebras open-sourced a complete local speech-to-speech pipeline chaining Nvidia Parakeet (speech recognition), Gemma 4 31B (the language model), and Qwen3-TTS (speech output). That pipeline already powers 9,000+ Reachy Mini robots — so local voice on Gemma 4 is shipping in real products, not just demos.

Related models

  • Gemma 2 9B — the previous mid-size Gemma; 12B Unified is its successor tier
  • Gemma 2 27B — previous large Gemma; compare with 26B-A4B and 31B
  • Gemma 3 270M — the tiniest Gemma, for microcontroller-class tasks
  • GPT-OSS — OpenAI's open-weight alternative in the prosumer tier
  • Best Ollama models — where the Gemma family sits in the wider local lineup
🎯
AI Learning Path

Go from reading about AI to building with AI

20 structured courses. Hands-on projects. Runs on your machine. Start free.

Or own it for life — Lifetime $149 $599, pay once
LM

Written by the Local AI Master Team

The team behind Local AI Master

We build Local AI Master around practical, testable local AI workflows: model selection, hardware planning, RAG systems, agents, and MLOps. The goal is to turn scattered tutorials into a structured learning path you can follow on your own hardware.

✓ Local AI Curriculum✓ Hands-On Projects✓ Open Source Contributor
More on Local AI Hardware
See the full AI Hardware Guide 2026 guide.
📚
Free · no account required

Grab the AI Starter Kit — career roadmap, cheat sheet, setup guide

No spam. Unsubscribe with one click.

🎯
AI Learning Path

Found your model? Now build something with it.

20 hands-on courses — RAG, agents, fine-tuning — all running locally. First chapter free, no card.

Or own it for life — Lifetime $149 $599, pay once
Free Tools & Calculators