AI Videos

What’s New in Gemma 4?

Google DeepMind unveiled Gemma 4, the latest open‑source family of large language models, now available under an Apache 2.0 license. The announcement emphasizes that the models can run directly on user‑owned hardware—smartphones, laptops, desktops, and even IoT devices—eliminating the need to send data to external servers. Gemma 4 spans a range of sizes: a 26‑billion‑parameter mixture‑of‑experts model with 3.8 billion active parameters for speed, a 31‑billion‑parameter dense model optimized for output quality, and highly efficient 2‑billion and 4‑billion‑parameter variants engineered for memory‑constrained environments. All models support a context window of up to 250 k tokens, native tool‑use, and multi‑step agentic workflows, while the smaller variants add real‑time audio‑vision capabilities and support for more than 140 languages. During the demo, the team showed the effective‑2B model answering a multilingual restaurant‑booking request and responding in English, illustrating its ability to understand and generate across languages. They also highlighted the model’s capacity to “see and hear the word,” underscoring the integrated multimodal perception. The release positions Gemma 4 as a trusted, enterprise‑ready foundation: it inherits DeepMind’s rigorous security vetting, offers on‑device inference for data privacy, and integrates with existing tooling. Companies can now deploy state‑of‑the‑art reasoning and coding pipelines without relinquishing control of proprietary data, accelerating AI adoption across regulated sectors.