Gemma 3

What Gemma 3 is

Gemma 3 is the third generation of the Gemma family of open models from Google, released on 12 March 2025 in four sizes: 1B, 4B, 12B and 27B parameters.^[1] Google described it as "the most capable model you can run on a single GPU or TPU", and said it was built from the research and technology behind Gemini 2.0.^[1] The technical report describes it as a multimodal addition to the Gemma family of open models, spanning 1 to 27 billion parameters.^[3]

The size range matters for how the models get used: the smallest variant is meant for environments where a full server is not available, while the largest is aimed at setups where one accelerator can still hold the model.

What it can do

Gemma 3 has a context window of 128,000 tokens, meaning a single prompt can carry a very long document or a long conversation before the model has to drop earlier material.^[1] The 4B, 12B and 27B models accept images as well as text, so they can be given a picture alongside a question; the 1B model works on text only.^[1]

The models ship with out-of-the-box support for more than 35 languages and pretrained support for more than 140, along with function calling and structured output, two features that let a model trigger external tools and return answers in a fixed format rather than free prose.^[1] Google also released official quantised versions, which store the model's numbers at lower precision to reduce its size and the compute needed to run it.^[1]

Reported results

Google reported preliminary human-preference evaluations on LMArena in which the 27B model reached an Elo score of 1338 and outperformed Llama3-405B, DeepSeek-V3 and o3-mini.^[1] Two caveats are worth keeping in mind when reading that number: the results are described as preliminary, and they measure how people rated pairs of answers, not performance on a specific task with a right answer. The comparison is also striking mainly because of scale — a 27B model is being placed against systems with far more parameters.

A change to the architecture

Long context is expensive in part because of the KV cache: when a model processes text, it stores key and value vectors for every token it has already seen so that it does not have to recompute them, and that store grows with the length of the context. The Gemma 3 technical report notes that the model's architecture was changed to reduce KV-cache memory, which tends to grow sharply as context gets longer.^[3] In practice this is what makes a 128,000-token window practical to serve rather than merely theoretically available.

Ecosystem and derivatives

Gemma 3 arrived on top of an existing community. Google said Gemma had passed 100 million downloads and more than 60,000 community variants in its first year.^[1] Derivatives built on the family included the 4B ShieldGemma 2, an image-safety checker, and MedGemma for medical applications.^[1][2]

That pattern — a base model released openly, then specialised versions built on it by both Google and outside developers — is the main reason the Gemma line matters. The base models provide general text and image capability; the derivatives take one of the smaller sizes and tune it for a narrow job such as flagging unsafe images or handling medical material.

Where it sits

Gemma 3 is best read as a step in an ongoing series rather than a one-off release. It keeps the family's emphasis on models that can be run and modified outside Google's own infrastructure, while adding capabilities that were previously associated with much larger systems: multimodal input, a long context window, broad language coverage, and tool use. Its stated comparison point is Gemini 2.0, whose research and technology it draws on, so the release also shows how Google has been pushing work from its closed models down into the open ones.^[1] The preliminary LMArena figures and the download counts are claims about momentum as much as about capability; what can be checked directly is the specification sheet — four sizes, a 128,000-token window, image support on the three larger models, and quantised builds for smaller hardware.^[1]