Gemma

Google's open weights, sized for hardware you own

Gemma is the open-weight counterpart to Gemini. Same provider, different proposition: you download these and run them.

Gemma 4 is the general model, taking text, audio and image input, with a context window between 128K and 256K. Two specialised variants sit beside it: EmbeddingGemma for retrieval, semantic similarity, classification and clustering, and ShieldGemma 2 for evaluating model inputs and outputs.

The licence is Apache 2.0, with prohibited-use and intended-use policies attached. That is a weaker constraint than Meta Llama's conditions and a stronger one than bare Apache, so read the policies rather than stopping at the licence name.

Multimodal input on a model you can run locally

Text, audio and image input on an open-weight model is the part worth noticing. Multimodal AI covers what media costs in a hosted request, and the calculation changes when the model is yours: no per-image token bill, and no audio leaving your network.

For a feature that processes user-supplied media under a privacy constraint, that combination is hard to get elsewhere.

Three variants, three different jobs

Treating Gemma as one model misses most of it.

Gemma 4 is the generation model, and the one to benchmark against your task.

EmbeddingGemma is a retrieval model, and it belongs in the comparison embedding models frames rather than being adopted because it shares a name. Sentence Transformers is how you would run it, and the embedding choice is stickier than a generation choice because changing it means re-embedding the corpus.

ShieldGemma 2 is a classifier for safety evaluation, which makes it the self-hosted alternative to a hosted moderation endpoint. Content moderation APIs covers what that tier is and is not for, and the argument holds whether the classifier is yours or someone else's.

Where it is meant to run

Locally, on mobile devices, through hosted services, or via the Gemini API.

The mobile line is the distinctive one. A model sized for a phone is a different product category from anything in the hosted comparison, and it opens on-device features where latency and privacy both matter and a network round trip is unacceptable.

For laptop use, Ollama and LM Studio are the usual ways to pull and serve it. For the browser, Transformers.js is the route, with the quantization caveats that come with it.

Community variants live on Kaggle Models and Hugging Face, so models on Hugging Face applies for telling a fine-tune apart from the base release.

Where it fits

Reach for Gemma when the model has to run on hardware you control and the size has to be modest: on a laptop, on a phone, or on a small server where a large open-weight model will not fit.

It sits beside Mistral's Ministral line and the smaller Qwen releases, and the choice between them is an evaluation rather than an argument. Run your own cases, as choosing the right model says, and be aware that at this size the gap between families on a specific task is often larger than any leaderboard suggests.

And running it yourself is still a commitment. Self-hosted models covers what that involves, which at these sizes is more tractable than for a frontier model and is not nothing.

Further reading

• Google, Gemma documentation: the variants, licence and where they run.
• Google Gemini: the hosted counterpart.
• Ollama and LM Studio: running it locally.
• Self-hosted models: what you take on.