Google Launches Gemma 4: Multimodal & Agentic Capabilities Now Available Under Apache 2.0 License

Google’s latest release, Gemma 4, is turning heads in the world of AI models. This family of open-weight models introduces a range of options including effective 2B and 4B edge variants, a 26B Mixture-of-Experts (MoE) model, and a substantial 31B dense model, all under the permissive Apache 2.0 license. This rollout not only marks a step forward in model capabilities but also enhances accessibility for developers and researchers alike.

One of the standout features of Gemma 4 is its native video and image processing capabilities, which are now integrated across the entire model lineup. The introduction of audio input for the smaller models also enriches the user experience, enabling more diverse applications. Context windows are expanded dramatically, reaching up to an impressive 256K tokens. Such a capacity allows the models to process large chunks of text, including extensive codebases and lengthy documents, all in a single prompt.

Google has shared that the 31B model achieves impressive benchmark results, scoring 84.3% on the GPQA Diamond and 80.0% on LiveCodeBench v6. These figures underscore the model’s substantial improvements over its predecessor, Gemma 3 IT’s 27B model, which scored only 42.4% on the same tests. Such advancements highlight enhanced capabilities in scientific reasoning and code generation, proving that Gemma 4 is not just a numerical upgrade but a significant leap forward in practical applications.

The architectural design of Gemma 4 plays a crucial role in its performance. The 26B MoE model employs a more flexible strategy, activating just 3.8 billion parameters during inference. This design enables rapid token processing while ensuring that the model remains efficient for various tasks. On the other hand, the 31B dense model is optimized for consistent performance, prioritizing ongoing reliability over peak parameter counts. Edge models tailored for mobile and IoT applications continue to impress by managing tight memory and power constraints while supporting a still-expanding context window.

Additionally, Google has integrated features that enhance the usability of these models for developers. The inclusion of native support for function-calling, structured JSON output, and direct system instructions means that users can create autonomous agents capable of sophisticated workflows. These capabilities are designed for fluid interactions with external tools and APIs, enabling developers to construct innovative applications with ease.

The performance benchmarks tell a compelling story as well. Estimates indicate that the 31B dense model achieved a LLMArena score of 1452, putting it in a performance category typically reserved for models with significantly larger parameter counts. This speaks volumes about the underlying technology and the efficiency of the design.

The response from the open-model community has primarily centered on usability and licensing. The Apache 2.0 license has drawn positive remarks from developers like Sam Witteveen. He praises the freedom it gives users to modify, fine-tune, and commercially deploy the model without excessive restrictions. Such freedom can stimulate innovation, as developers feel empowered to adapt the models to their unique needs.

This is an actual real Apache 2 license, which means for the first time, you can take Google’s best open model, modify it, fine-tune it, deploy it commercially, do whatever you want with it. No strings attached.

Moreover, industry experts like Nathan Lambert emphasize that Gemma 4’s true value will depend on its ease of use. He notes that slight variations in benchmark scores may become secondary if developers find the models easy to integrate into their existing workflows. With so many promising features and no licensing restrictions, many companies are likely to adopt Gemma 4 swiftly.

Gemma 4’s success is going to be entirely determined by ease of use, to a point where a 5-10% swing on benchmarks wouldn’t matter at all. It’s strong enough, small enough, with the right license, and from the U.S., so many companies are going to slot it in.

For developers eager to dive in, the distribution of Gemma 4 has been exceptionally broad. Weights are now available through platforms like Hugging Face and Kaggle, with convenient pathways through various tools including vLLM, llama.cpp, and NVIDIA NIM. Even more enticing, Kaggle is hosting the Gemma 4 Good Challenge, encouraging developers to create applications that promote meaningful positive change using these advanced models.