Gemma 4 on Google Cloud: Sovereign Deployment, Serverless GPUs, and What Enterprise Builders Actually Need

News

Gemma 4 on Google Cloud: Vertex AI, Cloud Run Blackwell GPUs, and Sovereign Deployment

Google Cloud now supports Gemma 4 across Vertex AI, Cloud Run with NVIDIA RTX PRO 6000 Blackwell GPUs (96GB vGPU), GKE, and Sovereign Cloud including air-gapped deployments. The Cloud Run path is particularly useful for builders who want serverless GPU inference that scales to zero with no reservations required.

Apache 2.0 on Gemma 4 Clears the Last Legal Blocker for Enterprise Forks and Sovereign Deployments

Gemma 4 is the first model in the Gemmaverse to ship under OSI-approved Apache 2.0, which removes the ambiguous usage restrictions that blocked enterprise procurement and government deployments on earlier versions. If you were waiting on legal clarity before building on Gemma, the blocker is gone.

Tools

GKE Agent Sandbox Pairs with Gemma 4 for Isolated Multi-Step LLM Code Execution

GKE now offers an Agent Sandbox that executes LLM-generated code in Kubernetes-native isolation with sub-second cold starts and up to 300 sandboxes per second. Combine that with GKE Inference Gateway cutting time-to-first-token by up to 70%, and you have a credible production stack for agentic workflows without custom infrastructure hacks.

LiteRT-LM Runs Gemma 4 E2B Under 1.5GB RAM with 4-Bit Weights and Tool Calling

LiteRT-LM now supports Gemma 4 E2B with under 1.5GB memory on some devices using 2-bit and 4-bit weights, constrained decoding for reliable tool calling, and 128K dynamic context on both CPU and GPU. On a Qualcomm Dragonwing IQ8, NPU acceleration hits 3,700 prefill tokens per second, which is fast enough to make on-device agents feel responsive.

Gemma 4 in Android Studio Agent Mode: Local Agentic Coding Without API Keys or Quotas

Android Studio now supports Gemma 4 as a local model for Agent Mode via Ollama or LM Studio, letting you run multi-file refactors, build error fixes, and feature generation entirely on your machine. The 26B MoE requires only 24GB total RAM and 17GB storage, which fits a mid-range developer workstation without a dedicated AI rig.

Analysis

Gemma 4 31B Dense and 26B MoE Benchmark Numbers: Where They Actually Stand

The 31B model ranks #3 among open models on Arena AI's text leaderboard while the 26B MoE activates only 3.8B parameters at inference, delivering fast token throughput on a single 80GB H100. The intelligence-per-parameter story is real but builders should verify benchmark tasks match their actual workloads before committing to serving infrastructure.