Gemma 4 Open Source Deep Dive: Apache 2.0, 140 Languages, and Running Local Agents at Scale

News

Gemma 4: Four Model Sizes, 400M Downloads, and the Most Capable Open Model Family Yet

The 31B dense model sits at #3 on Arena AI's open model leaderboard and beats models 20x its size. Four versatile sizes from E2B to 31B means you can match the model to your actual hardware budget without compromising on capability.

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

Cloud Run now supports Gemma-4-31B on NVIDIA RTX PRO 6000 Blackwell GPUs with 96GB vGPU memory and scale-to-zero pricing, which removes the cost argument against running large open models. The GKE Agent Sandbox adds Kubernetes-native isolation for LLM-generated code execution, which is the missing piece for production agentic deployments.

Analysis

Gemma 4 Apache 2.0 License Officially Confirmed: What It Actually Unlocks for Builders

Apache 2.0 is a meaningful upgrade from Google's previous Gemma terms because it gives developers unambiguous rights to modify, redistribute, and commercialize without navigating custom terms of service. This is the licensing clarity that enterprise and sovereign deployments have been waiting for.

Tools

LiteRT-LM Runs Gemma 4 E2B in Under 1.5GB RAM with Tool Calling and 128K Context

LiteRT-LM now supports constrained decoding, dynamic context lengths, and 2-bit/4-bit weights that let Gemma 4 E2B run on a Raspberry Pi 5 at 133 prefill tokens per second. The Python bindings and new CLI tool make it practical to wire this into IoT pipelines without deep framework expertise.

Gemma 4 on NVIDIA Platforms: From Jetson Orin Nano to Blackwell with NVFP4 Quantization

The 26B MoE model activates only 3.8B parameters during inference, making it fast enough for latency-sensitive edge workloads on a single H100. NVFP4 checkpoints for the 31B model are available now via NVIDIA Model Optimizer, cutting cost per token while keeping accuracy near 8-bit levels.