Gemma 4
Google's open-weights model family, in sizes from phone to workstation, free to run on your own device.
Quality
Modality
multimodal
Context
128K tokens (phone sizes) to 256K tokens
Access
open weights
Fabian's Take
"Gemma isn't one model, it's a whole series of releases in different sizes, and on Hugging Face you can get quantized versions that are much smaller still. Some even fit and run on a mobile phone. I find it very interesting for writing short texts, translating, and interpreting images, all without anything leaving the device."
Gemma 4 is Google’s family of open-weights models. Google released the first sizes on 2 April 2026 and has added more since; the 12B followed on 3 June. All of them are free to download and run under the Apache 2.0 license.
Which size for which device
The smallest sizes, E2B and E4B, are built for phones and other small devices. They read text, images and audio, with a 128K-token context. The 12B is the laptop size: Google says it needs about 16 GB of memory, and it also handles text, images and audio. For a workstation with a good graphics card there’s the 26B mixture-of-experts version, which only activates about 4 billion parameters per token and is fast for its size, and the 31B, the largest and strongest. Those two read text and images with a 256K-token context.
Quantized versions
On Hugging Face you’ll find quantized versions of every size: the same model with its numbers stored at lower precision, so it needs much less memory and runs faster, at a small cost in quality. That’s what makes the larger sizes usable on a laptop with plenty of memory, and the small ones on a phone. Tools like LM Studio and Ollama download and run them for you, and Google’s AI Edge Gallery app runs them on a phone.
Don’t expect cloud-model reasoning. For short texts, translations and describing images, where privacy matters, it’s a very interesting option for something that runs offline.
The Verdict
Best for: Short texts, translations and reading images on your own device, with nothing sent to the cloud.
Pros
- Runs entirely on your own device, so nothing leaves it
- Sizes for phones, laptops and workstations; quantized versions on Hugging Face are much smaller still
- Reads images, and the smaller sizes also take audio
- Apache 2.0 license
Cons
- Far behind the cloud models on complex reasoning
- Choosing a size and a quantized version takes some trial and error
Specs
- Pricing Free (runs locally)
- Cost Tier free
- ⚡Speed Tier fast
- License Apache 2.0