Google Gemma 4 Goes Apache 2.0: Open, Multimodal, and Phone-Ready

Google switches Gemma 4 to Apache 2.0

Google has released Gemma 4, the newest generation of its open model family. Built on research from Gemini 3, the series includes four variants—2B, 4B, 26B MoE, and 31B Dense—and supports image, video, and audio input. The most significant change is not the architecture, but the license: for the first time, Google is releasing Gemma under the Apache 2.0 license, replacing the more restrictive Gemma License that governed earlier releases.

What is in the lineup

The four-size structure is deliberate. The 2B and 4B variants are small enough to run on phones and edge devices; the 26B MoE and 31B Dense offer more capacity for server-side and agent workflows. All support up to a 256K context window, which is large enough for long documents, extended video transcripts, and multi-turn agent sessions. In public benchmarks, the 31B Dense model ranks third on the Arena AI leaderboard among open-source models, a credible position that suggests Google is not just releasing a research artifact but a model that can compete in real applications.

Why Apache 2.0 matters

Until now, the Gemma License imposed conditions that made many commercial builders nervous. Apache 2.0 removes that friction. It permits unrestricted commercial use, modification, and redistribution, provided the license and attribution are preserved. For startups and product teams, this means Gemma 4 can be embedded in paid products, shipped to customers, and forked for specialized domains without the legal uncertainty of a custom corporate license. That clarity is especially important for hardware vendors who want to pre-install a model on a device and for cloud providers who want to offer it as a managed service.

Multimodal input and on-device inference

The model’s ability to accept image, video, and audio in addition to text makes it a candidate for agents that perceive the world through device sensors rather than just chat windows. The 2B and 4B variants are positioned for local inference, which matters for latency, offline operation, and privacy. A user can ask a phone-based assistant to describe a scene, summarize a recorded lecture, or navigate an interface without sending every frame or utterance to a remote server. That architecture also reduces inference costs for application developers, because sensitive or repetitive workloads can stay on the device.

Agent workflows get a new baseline

Agentic systems need small, fast models that can run close to the user and larger models that can handle planning and reasoning. Gemma 4 spans both ends of that spectrum. The 31B Dense model can act as a reasoning backbone, while the 2B and 4B models can handle local perception and tool execution. The 256K context window supports long agent traces, allowing the model to retain conversation history, task instructions, and intermediate outputs across a session. These features align with the broader trend of agentic engineering becoming a distinct discipline, separate from prompt engineering and traditional AI engineering.

The open-source landscape

Gemma 4 enters a busy field. DeepSeek is expected to release V4 soon, likely as an open-source model, and Mistral is also building toward Mistral 4 while expanding into voice synthesis and developer platforms. Anthropic and OpenAI continue to push proprietary frontier models, with Anthropic emphasizing enterprise safety and OpenAI focusing on developer tools. Google’s move counters the narrative that open, competitive weights are the domain of smaller labs. By combining Apache 2.0 with multimodal support and a compact size tier, Google is making a direct bid to own the default open model for consumer devices and agent infrastructure.

Bottom line

Gemma 4 is not a single breakthrough model; it is a packaging decision. Google has aligned the license, the model sizes, and the modality set to make on-device, privacy-preserving AI agents practical for commercial builders. The 31B Dense model’s ranking shows competitive performance, while the 2B and 4B variants lower the barrier to running AI locally. Whether it becomes the standard will depend on tooling, downstream fine-tuning, and how well it runs on actual hardware. But on paper, it is one of the most useful open-source AI releases of the year.