Gemma 4 Context Windows, Multilingual Reach, and Production Serving: What Still Needs Your Attention

Analysis

Gemma 4's 256K Context and MoE Architecture: The Performance-Per-Parameter Story

The 26B MoE model activates only 3.8B parameters during inference, which means you get near-frontier reasoning at a fraction of the compute cost. For builders running local inference on consumer hardware, this is the model to benchmark first.

GKE Agent Sandbox and Inference Gateway: Running LLM-Generated Code Safely at Scale

GKE's new Agent Sandbox spins up 300 isolated Kubernetes environments per second for safe execution of LLM-generated code, and the Inference Gateway cuts time-to-first-token by up to 70 percent through capacity-aware routing. This is the infrastructure stack for anyone building multi-step agentic pipelines that need to run reliably in production.

Tools

Gemma 4 on Cloud Run with Blackwell GPUs: Serverless Inference at 96GB vRAM

Cloud Run now supports Gemma 4 31B on NVIDIA RTX PRO 6000 Blackwell GPUs with scale-to-zero, meaning you only pay when the model is actually running. This makes serverless local-quality inference viable for production workloads without managing GPU reservations.

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

LiteRT-LM now supports constrained decoding, dynamic context, and 2-bit/4-bit weights, letting Gemma 4 E2B run agent workflows on a Raspberry Pi 5 or Qualcomm Dragonwing IQ8. If you are building for IoT or embedded systems, this is the runtime to test.

NVFP4 Quantized Gemma 4 31B on Blackwell: 4-Bit Precision at Near 8-Bit Accuracy

NVIDIA's NVFP4 checkpoint for Gemma 4 31B cuts memory and improves throughput on Blackwell GPUs while keeping accuracy close to BF16. For anyone running high-volume inference, this is a meaningful cost-per-token reduction worth benchmarking.

News

Apache 2.0 on Gemma 4: What the License Change Actually Unlocks for Builders

Switching from Gemma's previous custom license to Apache 2.0 removes the legal ambiguity that blocked many commercial deployments. You can now modify, redistribute, and build products on Gemma 4 with the same confidence you would have with any OSI-approved model.