Google's Gemma 4 AI Models Go Truly Open Source, Ditching Restrictive License for Apache 2.0

Google is making a significant strategic pivot in its open AI model strategy, announcing the Gemma 4 family and, more critically, abandoning its custom, restrictive license in favor of the permissive Apache 2.0 license. This move directly addresses mounting developer frustration over the legal and usage limitations of previous Gemma releases, which offered 'open weights' but not true open-source freedom. The shift to a standard, well-understood license removes a major barrier to adoption and commercial use, signaling Google's intent to compete more aggressively in the developer-centric open model arena against rivals like Meta's Llama.

The new Gemma 4 suite arrives in four sizes optimized for local deployment, a core design principle of the Gemma line. The two largest variants—a 26B Mixture of Experts model and a 31B Dense model—are engineered to run unquantized on a single high-end 80GB Nvidia H100 GPU. While this hardware is expensive, it represents a 'local' machine for research labs and enterprises. For broader accessibility, Google notes these larger models can be quantized to lower precision to run on more consumer-grade GPUs, expanding their potential user base.

This licensing overhaul and hardware-focused release represent a clear attempt by Google to regain developer mindshare. By removing legal friction and emphasizing local, private deployment, Google positions Gemma 4 as a more flexible and trustworthy alternative to its cloud-bound Gemini models. The move increases competitive pressure in the open-weight model space, where licensing terms have become a key battleground for developer loyalty and downstream innovation.