Skip to content

How-ToRAM

How to Run Google's Gemma 4 Models Locally on Your PC

How to run Google's Gemma 4 models locally with Ollama: official tags and download sizes from the Ollama library on 14 November 2026, from the 4.6-7.5GB E2B build up to the 31B model, plus the commands ollama run gemma4 and gemma4:12b. Gemma 4 is officially Apache 2.0 and was announced by Google on 2 April 2026.

By , Editorial team

4 min read

How to Run Google's Gemma 4 Models Locally on Your PC

Gemma 4 is Google DeepMind's current family of open models, released on 2 April 2026 under an Apache 2.0 licence. Google says the models are built from the same research as Gemini 3, and that they are meant to run on your own hardware, not only in the cloud. The useful news for PC owners is that the models are on Ollama, LM Studio and Hugging Face, so you can chat with them offline once the files are on disk.

Forge PC has not benchmarked Gemma 4 or timed tokens per second on any graphics card. The sizes, commands and capability claims below come from Google's announcement and from Ollama's Gemma 4 library page, read on 14 November 2026.

Which Gemma 4 should you download?

Google's original announcement listed four sizes: Effective 2B (E2B), Effective 4B (E4B), a 26B mixture-of-experts model and a 31B dense model. Ollama also ships a 12B workstation build. On Ollama's library page, the download sizes and context windows are:

Ollama tag

What it is

Download size

Context

Inputs

gemma4:e2b

Edge model, 2.3B effective parameters

4.6–7.5GB

128K

Text, image, audio

gemma4 or gemma4:e4b

Edge model, 4.5B effective parameters (Ollama's default)

6.6–9.5GB

128K

Text, image, audio

gemma4:12b

Workstation dense model

7.7–8.0GB

256K

Text, image

gemma4:26b

Mixture of experts (about 3.8B active of 25.2B total)

16–19GB

256K

Text, image

gemma4:31b

Large dense model (30.7B)

19–20GB

256K

Text, image

The E2B and E4B builds are the ones Google describes as on-device models. They are the only sizes in Ollama's table that also accept audio. The 12B, 26B and 31B builds use a 256K context window. A 16GB graphics card is a comfortable home for the 12B download; the 26B and 31B packs need more memory or some of the model offloaded to system RAM, which is slower. Our local AI RAM guide and GPU for local models guide explain that trade-off, and our quantization explainer covers why these downloads are smaller than the raw parameter counts suggest.

Run it with Ollama

If you already use Ollama, this is a one-line job. If you do not, follow our Ollama or LM Studio guide first, then come back.

  1. Install Ollama from the official site and confirm it runs with ollama --version in a terminal.

  2. Start a chat with the default (E4B) model:

ollama run gemma4
  1. To pick a size explicitly, use the tag from the table, for example:

ollama run gemma4:12b

or gemma4:26b or gemma4:31b. Ollama will download the files on the first run, then reuse them.

  1. Type a prompt in the terminal to chat. To send an image, follow Ollama's usual multimodal syntax for that model, or use a front end that supports attachments. Ollama's model card says that for the best results with images or audio you should place that content before the text in the prompt.

  2. To use Gemma 4 from other apps, point them at Ollama's local API, which we describe in How to use Ollama's OpenAI-compatible API. Our best free local AI apps list has chat and coding front ends that can do this.

Ollama's documentation also describes a thinking mode that you can turn on by including a <|think|> token at the start of the system prompt, and says it already handles the chat template for you. Try the default first; only add that token if you want the model to show its chain of thought.

Other ways to run it

LM Studio. Google and DeepMind point people at Hugging Face and Ollama for downloads; LM Studio can load the same family if the model is in its catalogue or you import a GGUF. Our LM Studio model guide covers how to pick a file that fits your GPU.

llama.cpp. The llama.cpp project added Gemma 4 support earlier in 2026, including later work on audio input, and the project reached a 1.0 release in November; see our llama.cpp 1.0 news if you would rather run models from the command line.

Do not expect desktop-class results from the E2B model, and do not assume the 31B dense model will fit in 8GB of graphics memory. Start with gemma4:12b or the default E4B build, then step up if you have the VRAM and disk space.

Our verdict

The practical way to try Gemma 4 on a Windows PC is ollama run gemma4 for the small E4B default, or ollama run gemma4:12b if you have a mid-range GPU and want the 256K-context workstation build. Use gemma4:31b only if you have the memory for a 19–20GB download. The models are Apache 2.0, officially multimodal, and do not need a Google account once they are on your disk. Forge PC has not tested Gemma 4; this guide follows Google's announcement and Ollama's library page.

The bottom line

The practical way to try Gemma 4 on a Windows PC is `ollama run gemma4` for the small E4B default, or `ollama run gemma4:12b` if you have a mid-range GPU and want the 256K-context workstation build. Use `gemma4:31b` only if you have the memory for a 19–20GB download. The models are Apache 2.0, officially multimodal, and do not need a Google account once they are on your disk. Forge PC has not tested Gemma 4; this guide follows Google's announcement and Ollama's library page.

Sources

  1. 01Google, Gemma 4: Our most capable open models to date (2 April 2026)
  2. 02Google DeepMind, Gemma 4 model page
  3. 03Ollama, gemma4 library (model tags, sizes and documentation, read 14 November 2026)

About the author

Forge PC Staff

Editorial team

The shared byline for articles researched, written and edited by the Forge PC editorial team. How we source, check and correct our work is set out in our editorial policy.

More from Forge PC Staff

Read next

How to Enable AMD Fluid Motion Frames (AFMF 2.1) on a Radeon Graphics Card

How-To

How to Enable AMD Fluid Motion Frames (AFMF 2.1) on a Radeon Graphics Card

To enable AMD Fluid Motion Frames, install Adrenalin 25.3.1 or newer, then select the HYPR-RX profile (RX 7000 and newer) or turn on AMD Fluid Motion Frames under Gaming > Graphics, globally or per game. AFMF 2.1 works on RX 6000 cards and newer in DirectX 11, DirectX 12, Vulkan and OpenGL games, with VSync off; RX 6000 cards need exclusive full screen.

Forge PC Staff2 min

Two vs Four RAM Sticks for DDR5: Why 2x32GB Usually Beats 4x16GB

Versus

Two vs Four RAM Sticks for DDR5: Why 2x32GB Usually Beats 4x16GB

Two DDR5 sticks are better for most builds. AMD lists DDR5-5600 as the AM5 maximum with two modules but DDR5-3600 with four, and XMP or EXPO speeds above that are overclocking and not guaranteed. Buy 2x16GB or 2x32GB for 32GB or 64GB, and use four sticks, ideally a matched kit, only if you need 128GB or more.

Forge PC Staff3 min

AMD's FSR 4.1.2 Unreal Engine 5 Plugin Brings ML Upscaling to RX 6000 Cards

News

AMD's FSR 4.1.2 Unreal Engine 5 Plugin Brings ML Upscaling to RX 6000 Cards

AMD's FSR Unreal Engine 5 plugin now uses FSR Upscaling 4.1.2, bringing ML-based upscaling to Radeon RX 6000 cards and select Ryzen integrated graphics as well as RX 7000 and RX 9000, in UE 5.3 to 5.8 games that adopt it. It's DirectX 12 only for now and adds Frame Generation 4.0.2 with driver-assisted frame pacing on RX 9000 cards.

Forge PC Staff2 min

NVIDIA's DGX Spark HPC Developer Kit Links Four Sparks Into a 512GB Desk-Side Cluster

News

NVIDIA's DGX Spark HPC Developer Kit Links Four Sparks Into a 512GB Desk-Side Cluster

NVIDIA's DGX Spark HPC Developer Kit, announced at SC26 on 16 November 2026, links four DGX Spark systems in a ring over ConnectX-7 with no network switch, for 512GB of combined memory and 20TB of NVMe storage. NVIDIA says it can spread models of up to 700B parameters across the nodes with NVFP4. No UK kit price yet; Scan sells two-Spark bundles at about £12,845.

Forge PC Staff3 min