Gemma 4 released with 2.6B and 9B parameter variants

Gemma 4 hits 50 million downloads within weeks, fastest Google open model launch

Was integrated by a developer to replace an entire dash cam video analysis stack.

Community developer ported Gemma 4 to MLX-Swift, enabling local inference on Apple Silicon via LocallyAI app.

Gemma 4 model demonstrated self-terminating loop detection during a coding task, an emergent behavior for execution control.

Independent analysis declares Gemma4 models as best-in-class for small open LLMs.