Gemma 4 is four models, not one. Google released the family on April 3, 2026, under the Apache 2.0 license, the permissive commercial license that lets developers build products on top of the model without restriction. The four variants are E2B (Effective 2B), E4B (Effective 4B), 26B Mixture of Experts (MoE), and 31B Dense. The smaller two run on edge hardware. The larger two run on servers.
The numbers matter. The E2B and E4B models are designed for local execution on laptops and mobile phones, per Google’s technical documentation. The 26B MoE and 31B Dense variants support context windows up to 256K tokens, roughly 200,000 words of working memory per inference call. The E2B and E4B variants support up to 128K tokens, per Google’s documentation. HuggingFace has published model files for all four variants. The model also supports more than 140 languages.
Google designed Gemma 4 for advanced reasoning, agentic workflows, and code generation tasks. The company states the model accepts text, image, and video inputs, with audio listed as a supported modality, these are Google’s capability claims and have not been independently benchmarked in the 48-hour window since release. There are no Epoch AI evaluations or third-party benchmark results available yet. All performance framing should be read as vendor-stated until independent evaluations appear.
Why this matters for developers comes down to one constraint: compute. Running capable AI models locally has historically required either accepting weaker performance or paying for cloud inference. Apache 2.0 removes the licensing barrier. E2B and E4B remove the hardware barrier for teams already working with modern laptops and mobile chips. What remains is the capability question, and that answer will come from independent evaluations, not from Google’s release notes.
The context here is relevant. Open-source AI model releases have accelerated through early 2026, with multiple labs pushing capable open-weight models into the market. Google’s move with Gemma 4 keeps it competitive in the open-weight space while simultaneously seeding its developer ecosystem. Apache 2.0 is a deliberate choice, it’s the most commercially permissive license available, and it signals that Google wants Gemma 4 embedded in products, not just experimented with.
Watch for independent benchmark results over the next two to four weeks. The specific comparison points to track: how E2B and E4B perform against other edge-optimized open-weight models on standard reasoning and code tasks, and whether the multimodal input claims (particularly audio and video) hold up under third-party evaluation. Epoch AI typically indexes new frontier models within weeks of release. Until then, treat capability claims as directional, not definitive.
The Apache 2.0 license is the lead fact here, not the model sizes. Enterprise teams that have been waiting for an open-weight multimodal model they can legally embed in commercial products now have one from Google. Whether Gemma 4 is the best choice depends on independent evaluation results that don’t exist yet, but the legal and deployment flexibility is confirmed today.