Google has released Gemma 4 12B, an open-source artificial intelligence model designed to run entirely on standard enterprise laptops equipped with 16GB of VRAM or unified memory. The 11.95-billion-parameter model is available for download under the Apache 2.0 license on platforms including Hugging Face and Kaggle.

The model's key innovation is its "encoder-free" architecture that allows it to process raw audio waveforms and visual data directly without requiring separate encoding modules. This approach reduces both latency and memory consumption compared to traditional multimodal AI systems that use discrete encoders to translate audio and visual data before processing.

Gemma 4 12B can handle audio inputs up to 30 seconds and video processing for up to 60 seconds at one frame per second. The model features a 256,000-token context window, which enables it to process lengthy documents, and includes native function calling capabilities for autonomous agent applications.

The local operation capability addresses enterprise needs for data privacy and offline functionality. Organizations can process sensitive multimodal data on-premises without transmitting information to external APIs, which is particularly relevant for regulated industries such as healthcare, finance, and defense.

Google has made the model compatible with standard deployment frameworks including vLLM, SGLang, MLX, and llama.cpp. For Google Cloud users, the model can be deployed through the Gemini Enterprise Agent Platform Model Garden, Cloud Run, or Google Kubernetes Engine.

The release represents Google's continued focus on smaller, locally-executable AI models alongside the industry trend toward larger, more powerful systems that typically require cloud-based infrastructure.