Gemma 4 on TPUs, AMD, and Intel: Day-Zero Multi-Hardware Support Changes the Deployment Math

Tools

vLLM Ships Day-Zero Gemma 4 Support Across TPUs, AMD GPUs, and Intel XPUs

This is the first time vLLM has delivered Day 0 support on Google TPUs, AMD GPUs, and Intel XPUs simultaneously for a new model launch. For builders who want to avoid NVIDIA lock-in, this is a meaningful shift in what multi-hardware open model deployment actually looks like in practice.

LiteRT-LM Runs Gemma 4 E2B Under 1.5GB RAM with 128K Context on Raspberry Pi

LiteRT-LM's 2-bit and 4-bit weight support plus memory-mapped embeddings gets the E2B model running in under 1.5GB on some devices. The Qualcomm Dragonwing IQ8 NPU hits 3,700 prefill tokens per second, which makes agentic workflows on IoT hardware a real conversation.

News

Gemma 4 Official Launch: 31B Dense Ranks #3 Open Model Globally, Ships Apache 2.0

The 31B Dense model sits at #3 on Arena AI's open model leaderboard and the 26B MoE sits at #6, both beating models 20x their parameter count. Apache 2.0 licensing means no usage restrictions, no commercial fees, and no dependency on a cloud provider who can change pricing on you.

Android AICore Developer Preview Opens Gemma 4 Access Across Android Devices System-Wide

AICore lets developers tap into Android's built-in Gemma 4 model without shipping their own model weights, which cuts app size and removes backend infrastructure costs entirely. This is the clearest path yet to shipping on-device AI in a production Android app without per-request fees.

Analysis

Gemma 4 26B MoE Runs on a Single RTX 4070 with Only 8B Active Parameters During Inference

The MoE architecture activates only about 8B parameters per inference pass, which means a mid-range gaming GPU handles a 26B-class model without breaking a sweat. If you're doing high-volume text processing locally, the per-token cost after hardware is essentially electricity.