Google releases Gemma 3

What was released

On 12 March 2025, Google released Gemma 3, a family of models issued in four sizes: 1B, 4B, 12B and 27B ^[1]. The number in each name refers to the parameter count in billions — parameters being the weights a network learns during training — so the family runs from a small model suited to light tasks up to a larger one for heavier work.

What was new

Gemma 3 added vision input, meaning the models can take images as well as text ^[1]. It also came with a context window of 128,000 tokens ^[1]. A token is the unit a model reads and writes in, roughly a word or a piece of one, and the context window sets how much material the model can hold at once. A window of 128,000 tokens means long documents, sizeable code files or extended multi-turn conversations can be placed in front of the model in one go instead of being cut into pieces first.

Running on a single chip

Google said the models could run on a single GPU or TPU ^[1]. GPUs and TPUs are the accelerator chips used to train and serve neural networks; large models are normally spread across many of them, so aiming at one chip means a developer with workstation-class hardware can run the model locally rather than sending every request to a hosted service. Google also shipped official quantized versions, which store weights at lower numerical precision to reduce model size and compute requirements ^[1]. Quantization trades some numerical detail for a smaller footprint, which is what makes local use practical on modest hardware.

Languages and companions

According to Google, Gemma 3 supports more than 35 languages out of the box, with more than 140 languages covered in pretraining ^[1]. Out of the box means the released model can handle those languages without further training; pretraining coverage refers to languages the model saw while being trained on large text collections. Released at the same time was ShieldGemma 2, a 4B image-safety checker built on Gemma 3 ^[1]. A checker of this kind classifies images so that an application can screen what goes in or comes out, and building it on the same base model keeps it in step with the model it is guarding.

Where it sits

Gemma 3 is best understood as a release that puts several capabilities into one family at once: image input, a long context window, four sizes spanning a wide range of compute budgets, and official quantized builds for smaller footprints ^[1]. The single-GPU or single-TPU target is the part that shapes how the release gets used ^[1], because it decides whether a team can keep a model on its own machines. For someone choosing among the four sizes, the trade-off is the usual one: the 1B and 4B versions fit into tighter hardware and take less compute, while the 12B and 27B versions carry more capacity for harder tasks ^[1]. Shipping a companion safety model alongside the main family means an application that filters images does not have to source that filtering elsewhere ^[1].

What the release does not settle on its own is how well any of this works in a given setting; that depends on the task, the hardware and how the model is used. The concrete changes are the ones listed above — four sizes, image input, a 128,000-token window, single-accelerator operation, official quantized builds, broad language coverage, and a paired image-safety checker — all arriving together on 12 March 2025 ^[1].