Google DeepMind released Gemma 4 12B on Hugging Face on June 3, adding a 12-billion parameter model to its open-weights model family. The model is available for download and local deployment without requiring an API subscription.
Gemma 4 is the latest generation in Google's series of open models - models whose weights are publicly released so developers can run them on their own servers or hardware rather than routing requests through a cloud API. At 12 billion parameters (parameters are the numerical values the model adjusts during training; more generally correlates with greater reasoning capacity), this sits in the mid-size range: substantially more capable than compact models designed for on-device use, but manageable on server-grade hardware without requiring enterprise-scale GPU infrastructure.
The 12B size fills the gap between lightweight models that run cheaply but hit limits on complex tasks, and the largest open models that require significant resources to operate. For developers building local AI applications, fine-tuning custom models on proprietary data (training an existing model further to specialize it for a specific domain or task), or running private setups that avoid sending data to third-party APIs, Gemma 4 12B adds another viable option alongside existing open models from Meta, Mistral, and others.
Full technical specifications including benchmark performance, context window size (how much text the model can process in a single request), and licensing terms are available on the Hugging Face model page.