Gemma 4 on Google Cloud, TPUs, and the Full Production Stack: What's Ready Now

News

Gemma 4 Hits Vertex AI, Cloud Run, GKE, and Sovereign Cloud Simultaneously

Google pushed Gemma 4 across every major cloud deployment surface at once, including serverless Cloud Run with 96GB Blackwell GPUs and Sovereign Cloud for air-gapped environments. The 26B MoE variant is headed to Model Garden as a fully managed serverless option, which removes the infra overhead for most teams.

Tools

vLLM Ships Day 0 Support for Gemma 4 Across NVIDIA, AMD, Intel, and TPUs

vLLM's Day 0 multi-backend coverage for Gemma 4 means you can drop this into existing serving stacks without waiting for upstream support to catch up. AMD and Intel GPU support alongside TPUs makes this unusually portable for a same-day launch.

NVFP4 Quantized Gemma 4 Brings Near-8-bit Accuracy at 4-bit Cost on Blackwell

NVIDIA's NVFP4 checkpoint for Gemma 4 31B cuts cost-per-token while holding accuracy close to 8-bit, which is a meaningful efficiency unlock for anyone running inference at scale on Blackwell hardware. Jetson support from Orin Nano to Thor also extends the same model family from robotics to data center without retraining.

Gemma 4 in Android Studio Runs Full Agent Mode Locally with No API Key Required

Local Agent Mode in Android Studio with Gemma 4 handles multi-file refactors, build error loops, and feature scaffolding entirely on-device, which is a practical win for anyone working in regulated environments or just tired of quota limits. The 26B MoE runs on 24GB RAM, putting it within reach of most dev machines.

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

LiteRT-LM's 2-bit and 4-bit weight support gets Gemma 4 E2B under 1.5GB memory on some devices, and the new Python CLI makes it easy to prototype agentic pipelines on Raspberry Pi or Qualcomm edge hardware. The structured output and constrained decoding additions are what make tool-calling reliable enough to ship in production apps.