What Is Gemma? Deploying Google’s Open-Weight Models
Gemma is Google’s open-weight model family, distinct from the Gemini app. This article explains Gemma 4 specifications, effective parameters and memory, and why licensing must be checked by generation.
Meet this model
Gemma is Google’s family of open-weight generative models. Developers can download a particular release, run it on their own devices or services, and adapt it as appropriate. It is distinct from the Gemini chat app: downloading weights does not include the app’s search or file workspace. Choose both the model and inference framework, and check the license for the downloaded version. Gemma 4 and some earlier generations have different license files.
Different models target different hardware
The Gemma 4 overview available on October 20, 2026 lists E2B and E4B for mobile and edge devices, a 31B dense model, and a 26B A4B mixture-of-experts model. These are distinct models, not alternate display names. With constrained hardware, start with a release designed for that device instead of downloading the largest weights and then trying to compress them.
For offline document summaries, choose a local runtime and check extraction on a page containing exceptions. For images or audio, verify the media types accepted by that specific model. Support in one Gemma release does not establish the same multimodal configuration across every older model or parameter size.
Effective parameters are not the full weight size
The E in smaller Gemma 4 models means effective parameters. The overview explains that per-layer embeddings make static weight memory larger than that effective count may suggest. The 26B A4B model activates about four billion parameters per token but still loads all 26 billion for fast inference. Inferring total machine memory directly from “4B” misses important components.
Context-cache memory also grows with the prompt and response. Loading the weights does not guarantee that a long-document request will run. Quantized releases use different formats that must match the inference engine. When memory is exhausted, distinguish weight memory from context length rather than treating successful startup as support for arbitrarily long input.
Check Gemma 4 and earlier licenses separately
Google’s earlier Gemma Terms page explicitly limits its terms to models listed in the appendix and links separately to the Gemma 4 Apache 2.0 license. The old blanket description that Gemma uses one custom agreement is therefore insufficient. For distribution, use the files accompanying the target release and retain the required license and attribution information.
Open weights enable self-hosting and adaptation but do not automatically provide external knowledge, tool execution or correct business results. The Gemma model card notes that factual answers can be inaccurate or outdated. This article does not benchmark phones or computers. Its memory discussion comes from official specifications and makes no device-speed or document-accuracy promise.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.