Gemma 4 on vLLM, LiteRT-LM, and Google Cloud: The Full Stack Deployment Breakdown

Tools

Gemma 4 on vLLM: Day 0 Support Across NVIDIA, AMD, Intel, and Google TPUs

vLLM ships immediate Gemma 4 support across every major hardware backend including first-ever Day 0 TPU support, which is a big deal for teams running heterogeneous infra. If you are deploying Gemma 4 at scale, vLLM is now the fastest path to production across the widest hardware surface.

LiteRT-LM Runs Gemma 4 E2B Under 1.5GB RAM on Raspberry Pi and Qualcomm Edge Devices

LiteRT-LM gets Gemma 4 running in under 1.5GB memory with 2-bit and 4-bit weights, and hits 3,700 prefill tokens per second on a Qualcomm Dragonwing NPU. This is the most credible on-device agentic runtime story yet for IoT and robotics builders.

NVIDIA NVFP4 Quantized Checkpoints for Gemma 4 31B Cut Cost Per Token on Blackwell

NVIDIA ships NVFP4 quantized weights for Gemma 4 31B, delivering near-8-bit accuracy at 4-bit compute efficiency on Blackwell hardware. If you are optimizing inference cost on H100 or B100-class GPUs, this is the quantization path to benchmark first.

Android Studio Agent Mode Runs Gemma 4 Locally for Private, Offline Agentic Coding

Gemma 4 runs entirely on-device inside Android Studio Agent Mode, meaning your codebase never leaves your machine and you can run complex multi-file refactors without hitting any API quota. The 26B MoE variant requires only 24GB RAM, which makes it viable on a modern developer workstation.

News

Gemma 4 on Cloud Run with NVIDIA Blackwell RTX PRO 6000 GPUs and Scale-to-Zero Serving

Cloud Run now supports Gemma 4 31B inference on serverless Blackwell GPUs with 96GB vGPU memory and full scale-to-zero billing. For teams that want production-grade serving without managing GPU reservations, this is a viable path today.