Google DeepMind launches Gemma 4 open‑weight AI model series

Google DeepMind has released the Gemma 4 family of open‑weight large language models under an Apache 2.0 license. The series spans from tiny on‑device variants (E2B, E4B) to a 31‑billion‑parameter dense model and a 26‑billion‑parameter mixture‑of‑experts (MoE) version, all supporting multimodal reasoning, tool use, and a “thinking mode” for chain‑of‑thought processing.

Technical guides detail the architecture’s shared transformer core, efficiency tricks such as GQA and KV‑caching, and hardware requirements for local inference. Users can run the models on consumer GPUs via LM Studio, enabling the thinking channel by configuring reasoning parsing and prompt templates. Community testing of a quantized 12‑billion‑parameter Gemma‑4 agentic GGUF shows improved state‑keeping and tool‑calling accuracy compared with earlier releases, highlighting the advantage of local deployment for cost, latency, and privacy.

The releases aim to empower developers and researchers to experiment with multimodal AI, video and audio understanding, code generation from images, and function‑calling workflows without reliance on external APIs.