Google's Gemma 4 switches to Apache 2.0

Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Gemma 4 launched in four sizes — 26B Mixture-of-Experts, 31B Dense, and edge models E2B and E4B — all designed for local inference rather than cloud-only use, available on AI Studio, AI Edge Gallery, Hugging Face, Kaggle, and Ollama.
- Google switched Gemma's license from its custom, unilaterally updatable terms to Apache 2.0, a move the source frames as potentially the most consequential change, since the old license had a strict prohibited-use policy and could be read to extend to derivative AI models built with Gemma-generated synthetic data.
- The 26B MoE variant activates only 3.8 billion of its 26 billion parameters during inference, which Google says delivers substantially higher tokens-per-second than comparable models, while the 31B Dense variant prioritizes quality and is expected to be fine-tuned for specific use cases.
- E2B and E4B were co-developed with Google's Pixel team plus Qualcomm and MediaTek, targeting smartphones, Raspberry Pi, and Jetson Nano, and Google confirmed the next-gen Gemini Nano 4 will be based directly on these two edge models — the first official confirmation of a Nano 4.
- Gemma 4 adds native support for agentic workflows (function calling, structured JSON output, API instructions) and is tuned for code generation and visual-input tasks like OCR and chart understanding, with context windows of 128k tokens for edge models and 256k for the larger variants.
Why it matters: Developers who avoided Gemma over its custom license's prohibited-use clause and unilateral-update provision now get Apache 2.0's familiar, permissive terms, removing a major friction point for commercial adoption. The same release also seeds the next generation of on-device Pixel AI (Gemini Nano 4), meaning today's developer prototyping on E2B/E4B will carry forward to shipping phones — a direct pipeline from open model to consumer hardware that few open-weight families offer.
