Gemma 4

What it is

Gemma 4 is Google’s family of downloadable models. Its model card describes several sizes, with text and image inputs across the family and audio inputs on selected variants. The output modality is text. A model family name is not enough to identify what was actually used in an experiment.

How it works

Select a compatible artifact and runtime, then record the model identifier, quantization, runtime version and configured context. The model card’s maximum context does not mean a local application allocates that amount by default. Check the exact artifact and runtime support before assuming image, audio or tool-calling features are available.

Benefits

The family offers options for evaluating different resource and quality tradeoffs. Use the same reviewed examples across candidates and retain both failures and successful outputs. Include a question whose answer is absent from the documents to see whether the setup can decline to answer. Compare the resulting evidence with the requirements of the actual job.

Limits and risks

Gemma 4 uses Apache 2.0 licensing; older Gemma releases have different terms. Review the selected artifact and deployment requirements. Downloadable weights do not provide a hosted service, document retrieval pipeline or security boundary. Hardware needs, speed and answer quality depend on the variant and configuration. IonBin has reviewed the documentation but has not benchmarked these models.