Gemma 4 Recap: What the 31B Dense, 26B MoE, and Edge Models Mean for Your Local Stack

News

Gemma 4 Ships in Four Sizes: 31B Dense, 26B MoE, E4B, and E2B

The 31B is ranked #3 open model on Arena AI and fits on a single H100 in bfloat16, while the 26B MoE activates only 3.8B parameters at inference for fast local throughput. For builders, the practical takeaway is frontier-class reasoning on consumer GPUs with no API dependency. blog.google

Gemma 4 Is the First Gemmaverse Model Under Apache 2.0

Apache 2.0 is a meaningful upgrade from prior Gemma license terms because it gives builders clear rights to modify, redistribute, and commercialize without navigating bespoke terms of service. With 400M+ downloads and 100K variants in the community, the licensing clarity will accelerate derivative work fast. opensource.googleblog.com

Gemma 4 on Cloud Run with NVIDIA Blackwell GPUs Scales to Zero Between Requests

Cloud Run now supports Gemma 4 inference on RTX PRO 6000 Blackwell GPUs with 96GB vGPU memory, serverless scaling, and no reservations needed in us-central1 and europe-west4. For teams that want local model economics with cloud-scale burst capacity, this is the lowest-friction production path available right now. cloud.google.com

Tools

Gemma 4 Edge Models Run Offline on Phones, Raspberry Pi, and Jetson Nano

The E2B and E4B models were co-developed with Google Pixel, Qualcomm, and MediaTek and deliver near-zero latency on low-power hardware with native audio and a 128K context window. If you're building anything on-device or IoT, this is the most capable local option available today. 9to5google.com

Android Studio Agent Mode Now Runs Gemma 4 Fully Locally via Ollama or LM Studio

You can now run multi-file refactoring, build error resolution, and feature generation entirely on-device with no API key or internet connection required. The 26B MoE is the recommended size for most dev machines and needs 24GB total RAM, which is realistic on modern workstations. developer.android.com