Gemma.

Choose a model for your task

Gemma suits teams that want control over an AI feature's hosting and configuration. The family includes general language models and specialized variants. Choose a specific model with suitable inputs, licensing and hardware requirements, rather than a familiar name alone.

The Gemini consumer chat and downloaded Gemma weights are different ways to use Google technology. With an open model, you must set up the interface, request handling and result checking yourself. Having the weights does not automatically provide internet search or access to organizational documents.

Text and multimodal options

The official documentation covers question answering, summarization, reasoning and programming. Current variants support visual and other modalities to varying degrees. Check the selected model's card before configuration; do not assume the whole family's description applies to every size.

For file-based tasks, first determine how the application will extract data. PDF text, a page image and a table are different representations. The model receives what your pipeline prepares and can make mistakes if extraction is incomplete or poor.

Example: classify inquiries within your own environment

Suppose a company wants to classify short inquiries by topic: delivery, payment, returns and others. Write clear definitions and a small set of anonymized examples. Ask the model to return a category and brief explanation, allowing ‘insufficient data’ as an option.

Test ambiguous messages and inquiries covering two issues at once. A classification error must not automatically close an inquiry or change a customer's terms. This is a possible application design, not a built-in model feature or a test we conducted.

Memory, size and configuration

Google's documentation covers local execution, integrations and model configuration. Smaller configurations target limited resources, while larger ones require more memory. Reduced precision may help fit the weights, but assess results and speed in your own environment.

Account separately for input length, concurrent requests and memory for intermediate states. Successfully loading weights does not mean the system can handle a production workload. Start with a narrow use case and measure it against the device's actual limits.

Open-use conditions

Licenses can vary across Gemma generations and individual releases. The official page directs you to the relevant release's document; do not replace it with a general understanding of open source. Distribution, fine-tuning and running your own service require reading these terms.

Freely available weights do not remove hardware, storage and maintenance costs. Check model responses against source data; public applications need access controls, secure material handling and a clear error-correction procedure.

Suitable tasks

• Build an AI function on your own equipment
• Classify and summarize the text
• Set up a coding assistant
• Study the multimodal configuration of a suitable size