Google announces Gemma 4 with Apache 2.0 license and templates for local devices

Google announced this Thursday the Gemma 4 family, made up of new open source artificial intelligence models with available weights. The update represents the first major advancement in the line since the launch of Gemma 3 more than a year ago. Developers now have license Apache 2.0, which removes commercial restrictions present in previous versions.

The models support text, audio and image input, with context windows that reach 256 thousand tokens in the largest variants. Eles are primarily designed to run locally on affordable hardware, including consumer GPUs and mobile devices. The license change facilitates commercial use without additional obligations imposed by Google.

Technical improvements in reasoning and multimodal

The new models bring significant advances in reasoning, mathematics and following instructions when compared to the previous generation. Eles incorporate native support for function calling and JSON structured output generation, which benefits agentic workflows.

Code processing capability has been optimized for offline environments, achieving performance comparable to cloud services like Gemini Pro. Visual input support enables tasks such as optical character recognition and graph interpretation with greater accuracy.

• Variants include models Effective 2B and 4B optimized for low latency on smartphones.
• Collaboration with Qualcomm and MediaTek facilitates integration on mobile devices.
• Larger models run on a single 80GB H100 GPU without quantization.

Size variants and energy efficiency

The Gemma 4 family has four main size configurations. Versions 26B Mixture of Experts and 31B Dense offer high performance and run on server or workstation hardware. Já as Effective 2B and 4B prioritize efficiency for execution on edge devices.

The 26B MoE model activates just 3.8 billion parameters during inference, reducing latency and power consumption. Todas variants handle over 140 languages. Developers can download the full weights on platforms such as Hugging Face, Kaggle, and Ollama.

Immediate availability across platforms

The larger 31B and 26B models are available in AI Studio and Google. The lightweight E4B and E2B versions can be accessed in AI Edge Gallery. The full weights are available for immediate download from public repositories.

Companies and researchers can integrate the models into local applications without recurring API costs. Google also indicated that variants 2B and 4B will serve as the basis for the upcoming Gemini Nano 4 on Android devices.

Leia Tambem

Colby Minifie confirms Ashley Barrett’s powers in The Boys season five

Napoli x Milan in Serie A with confirmed lineups and where to watch live

New battery test puts Galaxy S26 Ultra ahead of iPhone 17 Pro Max in global ranking

Impact of switching to license Apache 2.0

The adoption of license Apache 2.0 eliminates the restrictions of the previous custom license, which included unilaterally updateable no-use policies. Desenvolvedores Gain greater control over data and business deployments.

This change should encourage the creation of new projects in the community, known informally as Gemmaverse. The focus on local execution reinforces the strategy of offering open alternatives to the closed models of the Gemini line.

Optimizations for specific hardware

The lightweight versions were developed in partnership with mobile chip manufacturers. Elas deliver near-zero latency in everyday tasks while maintaining reduced battery consumption. Testes indicate good performance on cards like Raspberry Pi and Jetson Nano.

Larger models maintain efficiency even in dense configuration or MoE. Reducing latency in local processing represents a practical gain for applications that require privacy and fast response without constant connection to servers.

Support multiple input modalities

In addition to text, the models process audio and images natively. Speech recognition improves over Gemma 3. Multimodal capability opens up possibilities for applications that combine different types of data in real time.

Developers can prototype agentic flows directly in AI Core Developer Preview using the lightweight variants. Essas implementations are forward-compatible with the future Gemini Nano 4.

The Gemma 4 family reinforces Google’s commitment to offering open models with accessible weights. The combination of improved performance, permissive licensing, and diverse hardware support expands options for those seeking locally runnable AI solutions.