Google released Gemma 4 on April 2, 2026, and the headline is not just that the models are more capable than anything in the Gemma family before. It is that they are designed to run locally, on hardware you already own, with no API calls, no usage costs, and no cloud dependency.

The new family ships in four variants: Effective 2B and Effective 4B for edge devices and phones, plus a 26B Mixture of Experts model and a 31B Dense model for higher-performance workloads. All four are released under Apache 2.0, which means commercial use is open with no restrictions.

The 26B and 31B models placed third and sixth on Arena AI’s text leaderboard, beating models twenty times their size.

What Gemma 4 Actually Does

The models support text, images, and audio as inputs. Context windows on the larger variants reach 256,000 tokens. They handle function calling natively, which means they can work as the reasoning layer inside agent workflows without needing a cloud model in the loop.

The smaller E2B and E4B variants run fully offline on phones, Raspberry Pi, and NVIDIA Jetson devices. Google developed these in collaboration with Qualcomm and MediaTek, optimising for the constraints of mobile silicon. The result is near-zero latency inference without a network connection.

For comparison, the models are up to four times faster than previous Gemma versions and use up to 60% less battery on mobile hardware.

Why Open-Source at This Level Changes the Calculation

The commercial AI model for the past three years has been: pay per token, per API call, or per seat. Your AI costs scale directly with how much you use it. That works fine at low volumes. It becomes a constraint at scale.

A locally deployed open-source model inverts that. Once it is running on your infrastructure, the marginal cost of additional inference is close to zero. That changes the economics of agentic AI completely. You can run high-frequency agent tasks, continuous document processing, or real-time voice transcription without watching a meter tick.

Gemma 4’s Apache 2.0 license makes this a realistic option for businesses, not just researchers. You can fine-tune it on your proprietary data, modify it for specific use cases, and deploy it in production environments without licence restrictions.

Where This Fits in the Competitive Picture

The local AI model space has been moving fast. Meta’s Llama 4 family, Mistral’s recent releases, and now Gemma 4 have all pushed capability into smaller, faster, more deployable packages. What Gemma 4 adds is Google’s investment in device-level optimisation, multimodal input natively, and a benchmark profile that competes with models that cost significantly more to run via API.

For enterprise teams, the practical implication is that the argument for self-hosted AI just got much stronger. The capability gap between cloud-hosted frontier models and self-hosted open models has been shrinking. Gemma 4’s placement on the Arena AI leaderboard suggests it has effectively closed for many common enterprise tasks.

The tools are already available across Hugging Face, NVIDIA NIM and NeMo, Ollama, Docker, and Google AI Studio. Teams do not need to wait for enterprise procurement cycles to start evaluating.

What This Means If You’re Building on AI

Gemma 4 is worth evaluating for any use case where you are currently paying per-token API costs at significant volume, or where data privacy requirements make cloud processing difficult. The combination of frontier-grade reasoning, multimodal input, and a permissive commercial licence covers a lot of ground.

It also matters for the voice AI and agent space specifically. Native audio input at this capability level, running locally, opens up on-premises voice AI deployments that were not commercially viable at this quality level six months ago.

If you want the playbook other teams are using with Claude and Codex right now, grab the free Working With Claude field guide. Download it here.