For two years, the most capable AI models lived behind paywalls and API meters. Google just made that harder to justify. Gemma 4, released April 2, is a family of four open-source models — ranging from 2 billion to 31 billion parameters — that handle text, images, video, and audio, run on consumer hardware, and ship under a fully permissive Apache 2.0 license with no usage restrictions.
For two years, the most capable AI models lived behind paywalls and API meters. Google just made that harder to justify. Gemma 4, released April 2, is a family of four open-source models — ranging from 2 billion to 31 billion parameters — that handle text, images, video, and audio, run on consumer hardware, and ship under a fully permissive Apache 2.0 license with no usage restrictions.
The license matters as much as the benchmarks. Previous Gemma releases carried custom terms that triggered legal review at large organizations, blocking adoption in healthcare, finance, and government. Apache 2.0 eliminates that friction entirely. The 31-billion-parameter flagship ranks third among all open models globally, while the 26-billion mixture-of-experts variant achieves nearly the same performance using only 4 billion active parameters — meaning it runs where the full model cannot. In a market where Meta's Llama, Alibaba's Qwen, and Mistral are all competing for the same developer base, Google is betting that removing every barrier — technical, legal, and financial — wins.
