Gemma 3
By Google
Gemma 3 is an open-weight, multimodal model family from Google that extends the Gemma line with image understanding, expanded multilingual support, and a longer context window, while remaining downloadable and self-hostable across a range…
Definition
Gemma 3 is an open-weight, multimodal model family from Google that extends the Gemma line with image understanding, expanded multilingual support, and a longer context window, while remaining downloadable and self-hostable across a range of parameter sizes. It adds image understanding to supported model sizes, expands the usable context window, and broadens multilingual coverage compared to Gemma 2, while remaining fully downloadable and self-hostable across a range of parameter sizes matched to different hardware budgets.
Overview
Gemma 3 builds on Gemma 2 by adding multimodal capability, allowing supported model sizes to accept image inputs alongside text, which brings the open-weight Gemma family closer to the multimodal capability of Google's closed Gemini models while keeping the weights openly downloadable for self-hosted use. Extending Gemma to accept image inputs required pairing the language model with a vision encoder similar in spirit to the approach used for Llama 3.2 and Gemini, so that visual information could be converted into a form the underlying language model could reason about jointly with accompanying text. The generation also expanded the model's context window considerably compared to earlier Gemma releases, enabling it to process longer documents and conversations in a single request, and it broadened multilingual coverage, improving quality across a wider range of languages than the previous generation supported. The expanded context window and broader multilingual coverage were both addressed through changes to training data composition and, for context length specifically, adjustments that let the model maintain coherence over longer input sequences than Gemma 2 could reliably handle, rather than through a change to the base parameter count alone. Gemma 3 was released across multiple parameter sizes, from smaller models suited to single-GPU or even CPU-based inference to larger variants offering stronger general and multimodal reasoning, continuing the family's emphasis on giving developers a choice that matches their available hardware rather than a single fixed model size. Gemma 3's addition of multimodal input is what most clearly separates it from Gemma 2 within the same open-weight family, while it remains distinct from Gemini in staying self-hostable, and it competes most directly with other multimodal open-weight releases like Llama 3.2 rather than with closed API-only models. As with Gemma 2, Gemma 3 is released under a license permitting broad commercial use, and it has been adopted for local inference applications, fine-tuning into specialized assistants, and integration into privacy-sensitive deployments where sending data to an external API is not desirable. Developers have used it for local applications that need to interpret both a document's text and any embedded images or charts, for multilingual products serving audiences beyond the languages well covered by Gemma 2, and for privacy-sensitive deployments where sending images to an external API is not acceptable. Gemma 3 represents Google's continued investment in a self-hostable open-weight lineup that tracks improvements from its closed Gemini research while giving developers full control over deployment, distinguishing its role from the API-only Gemini product line. Its multimodal capability is limited to supported model sizes rather than every variant in the release, and even the largest Gemma 3 model trails Google's closed Gemini flagship on the hardest multimodal reasoning tasks, so applications needing maximum capability rather than self-hosting control may still prefer a Gemini API.
Key Features
- Multimodal support for image inputs in supported model sizes
- Expanded context window compared to Gemma 2
- Broader multilingual coverage across supported languages
- Multiple parameter sizes from CPU-friendly to high-capability variants
- License permitting broad commercial and research use
- Fully self-hostable for offline and privacy-sensitive deployment