Gemma 4 / Download

Download Gemma 4 Open Models

Every checkpoint, every format, every source - with the exact repository IDs and pull commands. Apache 2.0, so you can download the weights, fine-tune them, ship them in a product and redistribute the result. New here? Start with what Gemma 4 can do.

Find my model All downloads Which format?

Links, tags and sizes checked on October 7, 2026.

Fastest start

# One command, any platform
ollama run gemma4:12b-it-qat

# 7.2 GB download · 256K context
# text, image and audio input

Smallest gemma4:e2b-it-qat4.3 GB
Balanced gemma4:12b-it-qat7.2 GB
Strongest gemma4:31b-it-qat19 GB

Pulled weights before mid-July 2026? Re-pull them - the refresh fixed the chat template and tool calling without changing the model names.

Find my model All downloads Formats Ollama tags Sources EmbeddingGemma 2 Related models Get it running FAQ

Find my model

Pick by the hardware you actually have

The most common mistake is downloading the biggest model your disk will hold rather than the biggest one your memory will run. Match to GPU memory or unified memory, not storage.

Bigger models need more memory - pick the largest one that fits.

1. How much memory?

2. How will you run it?

Suggestions are this site's guidance, based on Google's published memory figures plus room for the operating system and a working context. Full numbers are in the VRAM guide and technical specs; a longer questionnaire is in the runtime selector.

How to read the sizes. Google's memory figures cover loading the weights and already include about 20% overhead. The conversation itself needs more on top, and long contexts need a lot more - so leave headroom.

All downloads

Every Gemma 4 checkpoint

Repository IDs use capitals - gemma-4-12B-it, not 12b. Instruction-tuned (-it) is what you want unless you are doing your own post-training. What every suffix means is on the models page.

Five sizes, each with its own set of downloads. Sizes shown are the Ollama QAT builds.

31B 26B A4B 12B E4B E2B

Gemma 4 31B

Flagship

Dense · 256K context · text and image input · 30.7B parameters · benchmarks

Memory at 4-bit 17.5 GBMemory at 16-bit 69.9 GBOllama QAT download 19 GBLM Studio download 19.00 GB

gemma-4-31B-it-qat-q4_0-ggufRecommendedOfficial QAT GGUF - the best 4-bit build for llama.cpp, Ollama and LM Studio
gemma-4-31B-itInstruction-tuned, full precision Safetensors - for serving and fine-tuning
gemma-4-31BBase model, pre-trained only - for your own post-training
gemma-4-31B-it-assistantMTP drafter for faster decoding - a companion, not a chat model
gemma-4-31B-it-qat-q4_0-unquantizedQAT weights at half precision - for custom quantisation or MLX conversion
gemma-4-31B-it-qat-w4a16-ctCompressed tensors - for vLLM and SGLang serving

Ollama - QAT 4-bit, 19 GBollama pull gemma4:31b-it-qat

LM Studio - 19.00 GBlms get google/gemma-4-31b-qat

Hugging Face - GGUF filehf download google/gemma-4-31B-it-qat-q4_0-gguf

Hugging Face - full precisionhf download google/gemma-4-31B-it

Gemma 4 26B A4B

Mixture of Experts

About 3.8B parameters active per token · 256K context · text and image input

Memory at 4-bit 14.4 GBMemory at 16-bit 57.7 GBOllama QAT download 16 GBLM Studio download 15.60 GB

gemma-4-26B-A4B-it-qat-q4_0-ggufRecommendedOfficial QAT GGUF - the best 4-bit build for llama.cpp, Ollama and LM Studio
gemma-4-26B-A4B-itInstruction-tuned, full precision Safetensors - for serving and fine-tuning
gemma-4-26B-A4BBase model, pre-trained only - for your own post-training
gemma-4-26B-A4B-it-assistantMTP drafter for faster decoding - a companion, not a chat model
gemma-4-26B-A4B-it-qat-q4_0-unquantizedQAT weights at half precision - for custom quantisation or MLX conversion

Ollama - QAT 4-bit, 16 GBollama pull gemma4:26b-a4b-it-qat

LM Studio - 15.60 GBlms get google/gemma-4-26b-a4b-qat

Hugging Face - GGUF filehf download google/gemma-4-26B-A4B-it-qat-q4_0-gguf

Hugging Face - full precisionhf download google/gemma-4-26B-A4B-it

Gemma 4 12B

Unified

Encoder-free · 256K context · text, image and audio input · 11.95B parameters · Gemma 4 12B guide

Memory at 4-bit 6.7 GBMemory at 16-bit 26.7 GBOllama QAT download 7.2 GBLM Studio download 7.40 GB

gemma-4-12B-it-qat-q4_0-ggufRecommendedOfficial QAT GGUF - the best 4-bit build for llama.cpp, Ollama and LM Studio
gemma-4-12B-itInstruction-tuned, full precision Safetensors - for serving and fine-tuning
gemma-4-12BBase model, pre-trained only - for your own post-training
gemma-4-12B-it-assistantMTP drafter for faster decoding - a companion, not a chat model
gemma-4-12B-it-qat-q4_0-unquantizedQAT weights at half precision - for custom quantisation or MLX conversion
gemma-4-12B-it-qat-w4a16-ctCompressed tensors - for vLLM and SGLang serving

Ollama - QAT 4-bit, 7.2 GBollama pull gemma4:12b-it-qat

LM Studio - 7.40 GBlms get google/gemma-4-12b-qat

Hugging Face - GGUF filehf download google/gemma-4-12B-it-qat-q4_0-gguf

Hugging Face - full precisionhf download google/gemma-4-12B-it

Gemma 4 E4B

Edge

4.5B effective parameters · 128K context · text, image and audio input

Memory at 4-bit 4.5 GBMemory at 16-bit 17.9 GBOllama QAT download 6.1 GBLM Studio download 5.90 GB

gemma-4-E4B-it-qat-q4_0-ggufRecommendedOfficial QAT GGUF - the best 4-bit build for llama.cpp, Ollama and LM Studio
gemma-4-E4B-itInstruction-tuned, full precision Safetensors - for serving and fine-tuning
gemma-4-E4BBase model, pre-trained only - for your own post-training
gemma-4-E4B-it-assistantMTP drafter for faster decoding - a companion, not a chat model
gemma-4-E4B-it-qat-q4_0-unquantizedQAT weights at half precision - for custom quantisation or MLX conversion
gemma-4-E4B-it-qat-w4a16-ctCompressed tensors - for vLLM and SGLang serving
gemma-4-E4B-it-qat-mobile-transformersMobile-optimised QAT reference weights

Ollama - QAT 4-bit, 6.1 GBollama pull gemma4:e4b-it-qat

LM Studio - 5.90 GBlms get google/gemma-4-e4b-qat

Hugging Face - GGUF filehf download google/gemma-4-E4B-it-qat-q4_0-gguf

Hugging Face - full precisionhf download google/gemma-4-E4B-it

Phone builds: see mobile downloads and Google's LiteRT-LM.

Gemma 4 E2B

On-device

2.3B effective parameters · 128K context · text, image and audio input · mobile build from 1.1 GB

Memory at 4-bit 2.9 GBMemory at 16-bit 11.4 GBOllama QAT download 4.3 GBLM Studio download 4.20 GB

gemma-4-E2B-it-qat-q4_0-ggufRecommendedOfficial QAT GGUF - the best 4-bit build for llama.cpp, Ollama and LM Studio
gemma-4-E2B-itInstruction-tuned, full precision Safetensors - for serving and fine-tuning
gemma-4-E2BBase model, pre-trained only - for your own post-training
gemma-4-E2B-it-assistantMTP drafter for faster decoding - a companion, not a chat model
gemma-4-E2B-it-qat-q4_0-unquantizedQAT weights at half precision - for custom quantisation or MLX conversion
gemma-4-E2B-it-qat-w4a16-ctCompressed tensors - for vLLM and SGLang serving
gemma-4-E2B-it-qat-mobile-transformersMobile-optimised QAT reference weights

Ollama - QAT 4-bit, 4.3 GBollama pull gemma4:e2b-it-qat

LM Studio - 4.20 GBlms get google/gemma-4-e2b-qat

Hugging Face - GGUF filehf download google/gemma-4-E2B-it-qat-q4_0-gguf

Hugging Face - full precisionhf download google/gemma-4-E2B-it

Phone builds: see mobile downloads and Google's LiteRT-LM.

All repositories are published by Google on Hugging Face; the full lists are the Gemma 4, QAT Q4_0 and QAT mobile collections. The -it-assistant repositories are speculative-decoding drafters, not chat models with a persona.

Formats

Which file format do you need?

The same model in the wrong format is the most common reason a download does not run. Pick the row that matches your runtime.

Lower precision shrinks the same model to fit smaller hardware.

Format	Use it with	Precision	Notes
Safetensors	Transformers, vLLM, SGLang	16-bit	The reference format. Largest download, needed for fine-tuning.
QAT GGUF BEST 4-BIT	llama.cpp, Ollama, LM Studio	Q4_0	Quantisation-aware trained by Google, so it holds quality better than ordinary 4-bit. All five sizes. See the QAT guide.
QAT compressed tensors	vLLM, SGLang	4-bit weights, 16-bit activations	For high-throughput serving. Every size except 26B A4B.
QAT unquantized	Your own conversion tools	Half precision	QAT weights before quantisation, for custom schemes or MLX. All five sizes.
Community GGUF	llama.cpp, Ollama, LM Studio	Q2 to Q8	Many third-party quants exist. Prefer the official QAT build at 4-bit.
MLX	Apple silicon	4-bit or 16-bit	Available as `gemma4:*-mlx` tags on Ollama and through LM Studio. See Mac setup.
MXFP8 / NVFP4	Ollama on supported GPUs	8-bit / 4-bit floating point	Hardware-specific formats, offered as Ollama tags for every size.
LiteRT-LM and QAT mobile	Android, iOS, embedded	Mobile-optimised	On-device only, for E2B and E4B. See mobile downloads.

Don't go below 4-bit. Q2 and Q3 quants exist and will load, but quality on reasoning and code drops sharply - often below the next model size down at 4-bit. If 31B at 4-bit will not fit, run 26B A4B at 4-bit rather than 31B at 3-bit.

Ollama

Ollama tag reference

Download sizes as listed by Ollama. A range means the tag has more than one build for different hardware.

Pull once, then every app on your machine can use the model.

Tag	Download	Context	Notes
`gemma4:e2b`	4.6 - 7.5 GB	128K	On-device class
`gemma4:e2b-it-qat`	4.3 GB	128K	Smallest recommended
`gemma4:e4b`	6.6 - 9.5 GB	128K	Same build as `gemma4:latest`
`gemma4:e4b-it-qat`	6.1 GB	128K	Good laptop default
`gemma4:12b`	7.7 - 8.0 GB	256K	Unified multimodal
`gemma4:12b-it-qat`	7.2 GB	256K	Best quality per gigabyte
`gemma4:26b-a4b`	16 - 19 GB	256K	MoE, fast inference; `gemma4:26b` also works
`gemma4:26b-a4b-it-qat`	16 GB	256K	MoE at 4-bit
`gemma4:31b`	19 - 20 GB	256K	Flagship
`gemma4:31b-it-qat`	19 GB	256K	Best quality that fits 24 GB
`gemma4:31b-it-bf16`	63 GB	256K	Full precision
`gemma4:cloud`	-	256K	Hosted by Ollama, no local download

Every size also offers -it-q4_K_M, -it-q8_0, -it-bf16, -mlx, -mlx-bf16, -mxfp8 and -nvfp4 variants - see Ollama's full tag list. More on using it in the Ollama section.

Sources

Where to get it

All of these serve Google's weights. Pick whichever fits your workflow.

Different sources, the same official weights.

🤗

Hugging Face

The canonical source: every official checkpoint, plus community quantisations and fine-tunes.

Gemma 4 collection →

🦙

Ollama

One command, and it handles download and serving for you. The fastest path to a working local model.

gemma4 on Ollama →

💻

LM Studio

Desktop app with a built-in model browser. Search "gemma 4", click download, start chatting.

Gemma 4 in LM Studio →

📊

Kaggle

Google's own mirror, with free GPU notebooks if you want to try a model before downloading it.

Gemma 4 on Kaggle →

☁️

Vertex AI Model Garden

Deploy to a managed Google Cloud endpoint without downloading anything. Compare costs on the pricing page.

Open Model Garden →

🎛️

Google AI Studio

Try the 31B and 26B A4B in a browser and get an API key, if you would rather not host them at all. More places to chat.

Open AI Studio →

New · October 6, 2026

Download EmbeddingGemma 2

Google's open embedding model for search and RAG, built on the Gemma 4 architecture. It turns text, code, images, video and audio into vectors in one shared space - and at 1.3 GB or less, it is the smallest download on this page. Setup details are in Google's EmbeddingGemma 2 developer guide.

Hugging Face Ollama Kaggle

Different formats, same meaning, neighbouring points.

EmbeddingGemma 2

Embeddings

740M parameters, modular down to 270M · 768-dimension vectors · 8,192-token context · Apache 2.0

google/embeddinggemma-2OfficialWeights and model card on Hugging Face
embeddinggemma-2Ollama library - one command, four sizes
google/embeddinggemma-2 on KaggleSame weights, with free notebooks
EmbeddingGemma 2 collectionGoogle's release collection on Hugging Face
Developer guideGoogle's walkthrough: loading, cross-modal search, shorter vectors

Ollama - full model, 1.3 GBollama pull embeddinggemma-2

Ollama - text only, 378 MBollama pull embeddinggemma-2:270m

Hugging Facehf download google/embeddinggemma-2

Python librariespip install -U "sentence-transformers[image,audio,video]" transformers

Ollama tags

Tag	Download	Handles
`embeddinggemma-2:270m`	378 MB	Text and code
`embeddinggemma-2:440m`	714 MB	Text and images
`embeddinggemma-2:570m`	990 MB	Text and audio
`embeddinggemma-2:740m`	1.3 GB	Everything - same as `latest`

From Ollama's tag list. Each size also has -bf16, -mxfp8 and -nvfp4 variants.

Load only what you need

from sentence_transformers import SentenceTransformer

# Full model - 740M parameters
model = SentenceTransformer("google/embeddinggemma-2")

# Text only - 270M parameters
text_only = SentenceTransformer(
    "google/embeddinggemma-2",
    config_kwargs={"vision_config": None, "audio_config": None},
)

From Google's developer guide, which also lists LM Studio, LiteRT, MLX, vLLM and SGLang as supported.

It is not a chat model. EmbeddingGemma 2 produces vectors for search, retrieval and classification; pair it with a Gemma 4 model to write answers. Specs, task prompts, benchmarks and more code are in this site's EmbeddingGemma 2 guide.

Beyond the core models

Related Gemma models

If your workload is narrow, one of these will usually serve it better than a general model of the same size. Comparing generations or rivals? See Gemma 3, the model comparison and Gemma 4 vs Qwen 3.5.

Specialised models built around the same family.

New - Oct 6, 2026

EmbeddingGemma 2

Multimodal embeddings · 740M

Embeds text, code, images, video and audio into one shared space for search and RAG. Apache 2.0. It does not generate text.

Download EmbeddingGemma 2 ↑

Experimental - Jun 10, 2026

DiffusionGemma 26B A4B

Text diffusion

Generates text by refining blocks of tokens in parallel - roughly 4x faster output than autoregressive Gemma 4, with lower benchmark scores. Good at in-line editing and code infilling. Apache 2.0.

google/diffusiongemma-26B-A4B-it

Jan 15, 2026

TranslateGemma

Translation · 4B / 12B / 27B

Purpose-built for machine translation. Released before Gemma 4, so check its own licence terms and model card.

TranslateGemma on Kaggle

Jan 13, 2026

MedGemma 1.5

Medical · 4B

Tuned for medical imaging and clinical text. A research starting point, not a diagnostic tool - read the model card's limitations and terms first.

MedGemma at Google

FunctionGemma, ShieldGemma 2, T5Gemma v2 and the rest are listed with dates in Google's release history. Licences differ across these older models - only assume Apache 2.0 where the model card says so.

Before you download

Four things worth knowing

A quick check before a multi-gigabyte download.

📜

The licence is permissive

Gemma 4 is Apache 2.0: commercial use, modification, redistribution and proprietary derivatives are all permitted. Read the safety notes.

🔄

Re-pull old downloads

The July 2026 refresh changed the chat template and improved tool calling. Same names, different weights.

💾

Budget disk and bandwidth

Full-precision 31B is over 60 GB. Use hf download rather than git clone, which can store a second copy.

🔐

Prefer official repositories

Stick to google/ repos. Community re-uploads are usually fine but can be stale or mislabelled.

After downloading

Get it running

After the first download it runs with the network switched off.

# Install the CLI
pip install -U "huggingface_hub[cli]"

# Full precision checkpoint
hf download google/gemma-4-12B-it --local-dir ./gemma-4-12B-it

# Or just the official 4-bit GGUF - much smaller
hf download google/gemma-4-12B-it-qat-q4_0-gguf \
  --include "*.gguf" --local-dir ./gemma-gguf

# llama.cpp can pull straight from Hugging Face - no manual download
./build/bin/llama-server \
  -hf google/gemma-4-12B-it-qat-q4_0-gguf \
  --ctx-size 32768 \
  --n-gpu-layers 99 \
  --port 8080

# Ollama can run the same Hugging Face file directly
ollama run hf.co/google/gemma-4-12B-it-qat-q4_0-gguf:Q4_0

from transformers import pipeline
import torch

# Downloads on first run, then caches in ~/.cache/huggingface
pipe = pipeline(
    "text-generation",
    model="google/gemma-4-12B-it",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

out = pipe(
    [{"role": "user", "content": "Explain diffusion decoding briefly."}],
    max_new_tokens=400,
)
print(out[0]["generated_text"][-1]["content"])

# vLLM - OpenAI-compatible endpoint, production serving
pip install vllm

vllm serve google/gemma-4-26B-A4B-it \
  --max-model-len 262144 \
  --enable-auto-tool-choice \
  --tool-call-parser gemma4 \
  --reasoning-parser gemma4

# Ollama already serves on :11434 after a pull

Next steps: step-by-step download guide · install guide · how to use Gemma 4 · self-hosting · prompting · capabilities · Claude Code.

FAQ

Download questions

More answers on the full FAQ page, and known issues under limitations.

Short answers to the questions people ask most.

Is Gemma 4 free to download?

Yes. The weights are free under Apache 2.0, including for commercial use. You only pay for the hardware or hosting you run them on.

Which file should I download?

For running locally, the official QAT GGUF of the largest size your memory allows - or simply the matching -it-qat tag in Ollama. For serving or fine-tuning, the plain -it Safetensors repository.

How big is the download?

At 4-bit, from about 4.3 GB for E2B to about 19 GB for 31B through Ollama's QAT tags. Full-precision downloads are several times larger - 63 GB for the 31B on Ollama.

Why are download sizes bigger than the memory figures?

Google's memory figures are for a plain Q4_0 quantisation of the weights. Packaged downloads from Ollama and LM Studio include extra components and use different builds, so they come out somewhat larger and differ between sources.

What is the -it-assistant download?

A small draft model used alongside the main model to speed up generation. It is not a chat model and does nothing useful on its own. Details on the models page.

Can I download Gemma 4 on a phone?

Yes, the E2B and E4B sizes have mobile builds, starting at about 1.1 GB. See mobile downloads and the Android guide.

Do I need to download anything to try it?

No. You can try it in the playground or Google AI Studio, or call a hosted API.

Open weights. Yours to keep.

Apache 2.0 across the Gemma 4 family. Download once, run it anywhere, fine-tune it, ship it.

Browse all checkpoints Setup guide Official docs

Download Gemma 4 Open Models

Pick by the hardware you actually have

1. How much memory?

2. How will you run it?

Every Gemma 4 checkpoint

Gemma 4 31B

Gemma 4 26B A4B

Gemma 4 12B

Gemma 4 E4B

Gemma 4 E2B

Which file format do you need?

Ollama tag reference

Where to get it

Hugging Face

Ollama

LM Studio

Kaggle

Vertex AI Model Garden

Google AI Studio

Download EmbeddingGemma 2

EmbeddingGemma 2

Ollama tags

Load only what you need

Related Gemma models

Four things worth knowing

The licence is permissive

Re-pull old downloads

Budget disk and bandwidth

Prefer official repositories

Get it running

Download questions

Related guides

Open weights. Yours to keep.