Skip to content

Gemma models

Google DeepMind's open models, built from Gemini research. Gemma 4 runs from phones to one GPU, takes images and audio, and is now Apache 2.0.

  • Open weights
  • Models 6
  • Context up to 256K
  • Updated Nov 2026
Hugging Face Compare
On this page

At a glance

Type
Open-weight models
Developer
Google DeepMind (United States)
First release
Feb 2024
Models tracked
6 (3 current)
Longest context
256K tokens
Abilities
Tool callingReasoning modeImage inputAudio input
Licence
Gemma 4 is Apache 2.0. Gemma 3 and older use the Gemma Terms of Use, which allow commercial use but include a prohibited-use policy. MedGemma uses Health AI Developer Foundations terms.
Smallest local option
Gemma 3 270M · ~1 GB
Website
deepmind.google
Hugging Face
huggingface.co/google

Newest model

Gemma 4

Released Mar 2026 · 256K context

Smallest option

Gemma 3 270M

~1 GB VRAM · 292 MB download

Gemma models and versions

ModelReleasedSizesContextAbilitiesLicenceStatus
Gemma 4 Multimodal (image + audio), reasoning, 256K contextMar 2026 E2B · E4B · 12B · 26B · 31B E2B–31B (26B MoE, 4B active)256KToolsReasoningVisionAudioApache 2.0 Current
TranslateGemma Translation across 55 languagesJan 2026 4B · 12B · 27B 4B, 12B, 27B128KVisionGemma Terms of Use Current
MedGemma Medical text and image researchMay 2025 4B · 27B 4B, 27B128KVisionHealth AI Developer Foundations Current
Gemma 3n On-device model for phones and laptopsJun 2025 E2B · E4B E2B, E4B32KVisionGemma Terms of Use Previous generation
Gemma 3 Single-GPU multimodal model, 128K contextMar 2025 270M · 1B · 4B · 12B · 27B 270M–27B128KVisionGemma Terms of Use Previous generation
Gemma 2 8K context onlyJun 2024 2B · 9B · 27B 2B, 9B, 27B8KText onlyGemma Terms of Use Legacy

Sizes, context windows and abilities come from the Ollama library and Hugging Face model cards, checked 14 Nov 2026. Release months are the first public upload. Spotted a mistake? Tell us.

How much VRAM does Gemma need?

Download sizes per Ollama build and the GPU memory each needs with an 8K context. On a CPU-only server the same number is the RAM you need. Longer contexts need more memory; work it out with the LLM VRAM calculator.

Model4-bit (default)8-bit16-bitVRAM (4-bit)Fits onCommand
Gemma 4 E2B4.6 GB6 GB10 GB~6 GB8 GB GPU
Gemma 4 E4B6.6 GB9 GB16 GB~8 GB8 GB GPU
Gemma 4 12B8 GB13 GB24 GB~9.5 GB12 GB GPU
Gemma 4 26B19 GB28 GB52 GB~21.5 GB24 GB GPU
Gemma 4 31B20 GB34 GB64 GB~22.5 GB24 GB GPU
TranslateGemma 4B3.3 GB4.9 GB8.6 GB~4.5 GB8 GB GPU
TranslateGemma 12B8.1 GB13 GB24 GB~9.5 GB12 GB GPU
TranslateGemma 27B17 GB30 GB55 GB~19.5 GB24 GB GPU
MedGemma 4B3.3 GB5 GB8.6 GB~4.5 GB8 GB GPU
MedGemma 27B17 GB30 GB55 GB~19.5 GB24 GB GPU
Gemma 3n E2B7.8 GB6.9 GB11 GB~9.5 GB12 GB GPU
Gemma 3n E4B9.8 GB9.5 GB16 GB~11.5 GB12 GB GPU
Gemma 3 270M292 MB292 MB543 MB~1 GB8 GB GPU
Gemma 3 1B806 MB1.1 GB2 GB~1.5 GB8 GB GPU
Gemma 3 4B3.4 GB5 GB8.6 GB~4.5 GB8 GB GPU
Gemma 3 12B8.2 GB13 GB24 GB~10 GB12 GB GPU
Gemma 3 27B17 GB30 GB55 GB~19.5 GB24 GB GPU

VRAM = 4-bit download + about 10% runtime overhead + 0.5 GB for an 8K context, rounded up. Mixture-of-experts (MoE) models still need all their weights in memory, but they run much faster than a dense model of the same size, so they also work well split between GPU and system RAM.

Run Gemma 4 locally

# Install Ollama (Linux), then pull and chat
curl -fsSL https://ollama.com/install.sh | sh
ollama run gemma4:31b

New to Ollama? Follow Install Ollama on Ubuntu 26.04 (or 24.04) for GPU drivers, the systemd service and remote access.

What Gemma is good for

Laptop and edge AI

Gemma 4 E2B/E4B are designed for on-device use and take image and audio input.

Single-GPU multimodal assistant

Gemma 4 26B (MoE, 4B active) is fast on a 24 GB card; 31B is the strongest dense size.

Translation

TranslateGemma is tuned for 55 languages and keeps formatting.

Strengths and limits

Strengths

  • Gemma 4 moved to Apache 2.0.
  • Image input on every Gemma 4 size and audio input on Gemma 4 in Ollama.
  • Configurable thinking modes; 256K context on the larger sizes.

Watch out for

  • The "E" (effective) sizes download larger than their name suggests because they include vision and audio encoders.
  • Gemma 3 and older keep the more restrictive Gemma Terms of Use.

About Gemma

Gemma is Google DeepMind’s family of open models, built from the same research as Gemini. Gemma 1 (February 2024) and Gemma 2 (June 2024) were text-only with an 8K context. Gemma 3 (March 2025) added image input and a 128K context, with sizes from 270M to 27B, and became one of the most popular single-GPU models.

Gemma 4 (March 2026) is the current generation. Every size reasons with configurable thinking, takes images, and the Ollama builds also accept audio. The line-up mixes dense and mixture-of-experts models: E2B and E4B for devices, 12B and 31B dense, and a 26B MoE that activates about 4B parameters per token. Gemma 4 is also the first Gemma released under Apache 2.0.

Google also publishes specialised variants: TranslateGemma for translation, MedGemma for medical research, and EmbeddingGemma for search (see embedding models).

Guides for running Gemma

Tool LLM VRAM Calculator: How Much GPU Memory Do You Need? Updated Nov 2026 AI & LLMs How to Install Ollama on Ubuntu 26.04 and Run LLMs Locally Updated Nov 2026 AI & LLMs How to Install llama.cpp on Ubuntu 26.04 (CPU and NVIDIA GPU) Updated Nov 2026 AI & LLMs How to Install vLLM on Ubuntu 26.04 (OpenAI-Compatible LLM Server) Updated Nov 2026 AI & LLMs How to Install Open WebUI on Ubuntu 26.04 (Self-Hosted ChatGPT Alternative) Updated Nov 2026

Frequently asked questions

What is the latest Gemma model?

The newest Gemma model is Gemma 4, released in Mar 2026. It has a 256K-token context window and supports tool calling, reasoning mode, image input, audio input.

How much VRAM do I need to run Gemma locally?

The smallest option, Gemma 3 270M, is a 292 MB download at 4-bit and needs about 1 GB of GPU memory (or RAM on a CPU-only machine) with an 8K context. Larger sizes need more; see the hardware table on this page, and use the VRAM calculator for longer contexts.

How do I run Gemma with Ollama?

Install Ollama, then run "ollama run gemma3:270m". Ollama downloads the 4-bit build and starts a chat; the same model is also available on the local API at http://localhost:11434.

Can I use Gemma commercially?

Gemma 4 is Apache 2.0. Gemma 3 and older use the Gemma Terms of Use, which allow commercial use but include a prohibited-use policy. MedGemma uses Health AI Developer Foundations terms. Always read the licence of the exact model you deploy; it is linked in the models table.

What does the "E" in Gemma E2B and E4B mean?

"Effective" parameters. These models use per-layer embeddings that can be kept out of accelerator memory, so they behave like 2B or 4B models while the files are larger.

Compare Gemma

← All AI models

Back to top