Google DeepMind made Gemma 4 available on April 2, 2026. Four model variants span the full deployment spectrum: E2B and E4B for mobile and edge devices, a 26B Mixture of Experts model, and a 31B Dense model for datacenter-class inference. Every variant ships under an Apache 2.0 license, a meaningful change from prior Gemma licensing terms that carries direct implications for enterprise adoption.
The architecture comes directly from Gemini 3 research. According to The Register, all four variants support native function calling and more than 140 languages. Multimodal inputs, text, image, and video, are supported across the family, with audio support available in select variants, per Google DeepMind’s official release. The models are purpose-built for agentic workflows and local execution on diverse hardware.
The 31B Dense model ranked third among open models on the Arena AI text leaderboard at the time of release, according to available reporting, a community evaluation that reflects real-time voting rather than a static benchmark, and an independent evaluation from Epoch AI has not yet been published. The 26B Mixture of Experts model is reported to activate approximately 3.8 billion parameters during inference, per The Register’s coverage. Reported context windows reach up to 128K tokens for smaller variants and 256K tokens for the larger models, though these figures come from journalism rather than the official source and should be treated as approximate.
Google and NVIDIA collaborated on GPU optimization. NVIDIA’s official blog confirms that Gemma 4 is optimized for RTX-powered PCs, DGX Spark, and Jetson Orin Nano, making the same model family deployable from a laptop to a full datacenter node. Modular, a day-zero launch partner, reports approximately 15% higher throughput on NVIDIA B200 hardware compared to vLLM, a partner benchmark, not an independent one. The Gemma 4 family is available on Hugging Face and the Modular platform at launch.
The Apache 2.0 license is the story within the story. Prior open-weight releases have carried more restrictive terms that limited commercial use, creating friction for enterprises that wanted the cost advantages of local inference without the compliance risk of unclear licensing. Apache 2.0 removes that friction. An enterprise can deploy Gemma 4 in production without negotiating terms or tracking use-case restrictions. That alone makes this release strategically significant regardless of how the benchmarks eventually settle.
The competitive context matters too. The Register’s framing, “Google battles Chinese open-weights models with Gemma 4”, points to something real. Strong open-weight releases from Chinese labs have pulled developer attention and enterprise pilots away from US-based models. A capable, permissively licensed model from the same research base as Gemini 3 is a direct competitive response. Whether it succeeds depends on independent evaluation that isn’t available yet.
Watch for Epoch AI’s independent benchmark results. When they publish, the Arena AI community ranking and the Modular throughput figures will have a proper reference point. Watch also for enterprise adoption signals, specifically, whether the Apache 2.0 terms and the Gemma 4 E2B/E4B variants show up in edge deployment announcements over the next 60-90 days. That’s the real signal on whether Google has closed the gap with Chinese open-weight alternatives.
The Gemma 4 release is best understood as three bets at once: a bet on Apache 2.0 licensing as a competitive differentiator, a bet on Gemini 3 architecture translating to open-weight performance, and a bet on agentic and edge deployment as the next battleground. All three are reasonable bets. None are proven yet.