Google DeepMind Releases Gemma 4: Four Open Models with Multimodal, 256K Context, and Agentic Capabilities Under Apache 2.0

Google Blog·high signal

Google DeepMind released Gemma 4 on April 2 with four model sizes (E2B, E4B, 26B MoE, 31B Dense) under Apache 2.0 featuring multimodal capabilities (text, vision, audio), 256K context windows, and native thinking/tool-calling optimized for agentic workflows. Day-0 ecosystem support shipped across vLLM, llama.cpp, Ollama, and Unsloth, with the 26B MoE variant hitting 162 tok/s decode on RTX 4090 at 19.5GB VRAM and 34 tok/s on Mac mini M4 — the strongest open-weight agentic model family yet for local deployment.