One model replaces three pipelines
Google released Gemma 4 12B on June 3, 2026. 12 billion parameters, encoder-free multimodal across text, image, and audio. Runs on 16GB laptops at 4-bit quantization using roughly 8GB of memory. Ships under Apache 2.0: no restrictive license agreements, no vendor approval gates, no usage restrictions for defense deployment.
Before Gemma 4 12B, an edge inference stack needed separate models for each modality. Object detection required a vision model. Audio transcription required a speech model. Reasoning over combined inputs required a language model with adapter layers. Each model consumed memory, required its own pipeline, and added latency. A three-model stack on a 16GB device left almost no headroom for the application layer.
Gemma 4 12B collapses that into a single model. The vision embedder uses 35 million parameters to process 48x48 pixel patches via a single matrix multiply. Audio input arrives as raw 16kHz waveform sliced into 40ms frames and projected linearly into the embedding space. No separate encoder, no handoff between models, no pipeline orchestration.