Gemma 4 on NVIDIA RTX and DGX Spark: Local Agentic AI Gets a Hardware Upgrade

Tools

Gemma 4 Optimized for NVIDIA RTX, DGX Spark, and Jetson Orin Nano

Google and NVIDIA co-optimized all four Gemma 4 variants for NVIDIA hardware, from Jetson edge modules to RTX workstations to DGX Spark. If you're running local agents, this is the clearest hardware deployment path available today with day-one Ollama and llama.cpp support.

Android Studio Gets Gemma 4 Local Agent Mode with Full Offline Code Execution

Gemma 4 now powers Android Studio's Agent Mode locally, meaning your code never leaves your machine and you skip API quotas entirely. The 26B MoE runs on 24GB RAM, which is actually achievable on a developer workstation, making this the most practical offline coding agent setup for Android builders so far.

LiteRT-LM Runs Gemma 4 E2B Under 1.5GB RAM with Tool Calling on Raspberry Pi and Mobile

LiteRT-LM processes 4,000 input tokens across two agent skills in under 3 seconds on-device, and runs Gemma 4 E2B on Raspberry Pi 5 CPU at 133 prefill tokens per second. Constrained decoding and dynamic context support make this the most production-ready edge inference stack for agentic workflows right now.

News

Gemma 4 Family Overview: Four Sizes, Apache 2.0, 256K Context, and Agentic-First Design

The full Gemma 4 lineup ships with native function calling, structured JSON output, 140-plus languages, and up to 256K context under Apache 2.0. The 31B model ranks third among open models on Arena AI, beating models 20x its size, which changes the cost math for anyone self-hosting.

Gemma 4 on Google Cloud: Serverless Blackwell GPUs, GKE Agent Sandboxes, and Sovereign Deployment

Cloud Run now supports Gemma 4 on NVIDIA RTX PRO 6000 Blackwell GPUs with 96GB vGPU memory, scale-to-zero, and no reservations needed. The GKE Agent Sandbox addition is the detail worth tracking here as it lets you safely execute LLM-generated tool calls in isolated Kubernetes environments with sub-second cold starts.