Google is rolling out Gemma 4 across Google Cloud, and the pitch is pretty simple: this is Google’s most capable open model family so far, now wired directly into the cloud products developers already use every day — Vertex AI, Cloud Run, GKE, TPUs, and Sovereign Cloud. It’s built on the same research stack as Gemini 3, but unlike Google’s proprietary models, Gemma 4 ships with open weights under a standard Apache 2.0 license, which is a big deal for anyone who wants maximum freedom to ship real products without legal headaches.

At a high level, Gemma 4 comes in four sizes: Effective 2B (E2B), Effective 4B (E4B), a 26B Mixture of Experts model, and a 31B dense model. The smaller E2B and E4B variants are tuned for edge and on-device scenarios — think phones, browsers, small servers — while the 26B MoE and 31B dense models are aimed at heavier enterprise workloads where you care about reasoning quality, long context, and throughput. Context windows go up to 256K tokens on the larger models, with multimodal inputs covering text plus vision and audio, and even video support at the high end, so the models can chew through big codebases, long documents, logs, or media-heavy workloads in one go.

The other headline move is licensing. Earlier Gemma generations had custom terms that made some enterprises nervous, especially around sensitive or regulated deployments. Gemma 4 switches to Apache 2.0, the same license used by many mainstream open-source projects, which effectively removes that friction: you can fine-tune, embed, and ship Gemma 4 models inside commercial products without special carve‑outs, while still keeping them in your own infrastructure if you want. That’s why you’re also seeing Gemma 4 pop up beyond Google Cloud — it’s already on Hugging Face, Kaggle, and Ollama, plus Google’s own AI Studio and AI Edge Gallery.

On Google Cloud itself, Vertex AI is the most straightforward starting point. You can pull Gemma 4 from Model Garden and deploy it to your own managed endpoints, picking the compute profile that matches your workload and cost envelope. For teams that need differentiation, Vertex AI Training Clusters let you fine‑tune Gemma 4, with recipes optimized for SFT and large‑scale training, and support for NVIDIA NeMo Megatron, so you can push from the small E2B edge model all the way up to the 31B dense variant. Google is also rolling out a fully managed, serverless option for the 26B MoE model in Model Garden, so you don’t even have to think about infrastructure but still get a high‑throughput, relatively low‑latency model for production.

If you’re building AI agents rather than just single-turn prompts, Gemma 4 is clearly designed with that in mind. The models focus on reasoning, multi‑step planning, structured outputs, and function calling, and Google is pairing that with its Agent Development Kit (ADK), an open‑source framework for wiring up tools, memory, and workflows. ADK lets you plug Gemma 4 into agents that call APIs, run code, or orchestrate multi‑step tasks, with Gemma 4 providing the brain and ADK handling the plumbing around it.

Cloud Run is the “I want GPUs without managing GPUs” option. With support for NVIDIA RTX PRO 6000 (Blackwell) GPUs and 96GB of vGPU memory per instance, you can run something as heavy as Gemma-4-31B-it on fully managed, serverless GPUs. Cloud Run handles auto‑scaling for you, including scaling to zero when idle, and you can tune CPU and memory per container to match your inference profile, which keeps costs under control while still reacting quickly to traffic spikes. Google is also publishing hands‑on codelabs showing how to deploy Gemma 4 with vLLM on Cloud Run, making it more approachable for non‑infra‑experts.

For teams that want deeper control, GKE is where things get interesting. You can deploy Gemma 4 on Kubernetes with your choice of GPUs or TPUs, custom autoscaling policies, and integration into your existing microservices stack. Google is leaning heavily on vLLM as the serving layer here, so you can scale from zero to peak traffic while making good use of KV‑cache and memory, and you get a more “cloud-native” LLM deployment story instead of a one‑off box of GPUs in the corner. On top of that, the GKE Inference Gateway adds latency‑aware routing: it watches real‑time accelerator metrics and uses predictive scheduling to send each request to the server that can respond fastest, which Google says can cut time-to-first-token by up to 70% in some cases when paired with features like predicted-latency-based scheduling in llm-d.

Gemma 4 is also being pushed hard on TPUs. Across GKE, Compute Engine (GCE), and Vertex AI, you can serve, pretrain, and post‑train the 31B dense and 26B A4B MoE variants using open‑source stacks like MaxText for training and vLLM TPU for serving. MaxText gives you recipes for post‑training targeted tasks like text analysis, code reasoning, or image understanding, and vLLM TPU provides high‑throughput serving on Google’s accelerator fleet with prebuilt containers and quickstart tutorials. For teams that have standardized on TPUs or want to squeeze maximum performance out of Google’s hardware, this is the path that lines up with Google’s own internal best practices.

One of the more strategic angles in this launch is Sovereign Cloud. Gemma 4 is rolling out across Google’s various sovereignty offerings — from data‑bounded public cloud regions to dedicated environments like S3NS in France, all the way to air‑gapped and on‑prem setups via Google Distributed Cloud. Because the models are open‑weights, enterprises and governments can deploy Gemma 4 in tightly controlled environments, keep all data and logs within national borders, manage their own keys and encryption, and still fine-tune for local languages, regulations, or domain‑specific tasks. For regulated industries and public‑sector buyers, that mix of open weights plus sovereignty and compliance is the main selling point versus pure SaaS models.

Zooming back out, what Google is doing with Gemma 4 on Cloud is essentially filling in the “open yet enterprise-grade” gap in its AI lineup. You get a model family that covers edge devices through to big server deployments, strong reasoning and multimodal capabilities, long context, and a permissive license — all tied into the managed infrastructure, agents framework, and sovereignty story of Google Cloud. For developers and companies choosing stack today, it means you can start small with a 2B or 4B model, experiment in Vertex AI or Cloud Run, and later graduate to highly optimized GKE or TPU setups — without having to switch model families or rewrite your entire app.