Google introduces Gemma 4 12B, an encoder‑free multimodal model

Google has launched Gemma 4 12B, a 12‑billion‑parameter multimodal model that forgoes traditional encoders in favor of a unified architecture [Google Blog]. The model accepts text, images and audio through a single processing pipeline, eliminating the need for separate encoder modules.

── What shipped ──

Gemma 4 12B builds on Google’s prior multimodal work but replaces the encoder stack with a shared transformer backbone. Google says the encoder‑free design reduces computational overhead and improves throughput for mixed‑modality workloads [Google Blog]. The model is available via the company’s AI platform, with pretrained weights and a developer‑friendly API.

── Why it matters ──

The unified approach simplifies system integration: developers no longer need to stitch together distinct text, vision and speech encoders, which can lower engineering effort and hardware costs. By cutting the number of separate components, the architecture also reduces latency in real‑time applications such as interactive assistants and multimodal search. Finally, the model’s 12‑billion‑parameter scale positions it competitively with other large‑language‑plus‑vision systems, offering a single model that can be fine‑tuned for diverse tasks.