Google Open-Sources Gemma 4 12B: Encoder-Free Multimodal for 16GB VRAM

Google released Gemma 4 12B, a dense multimodal model running on 16GB VRAM without encoders.

💡 What You Will Learn

Google released Gemma 4 12B, a dense multimodal model running on 16GB VRAM without encoders.

Today, Google introduced a brand-new member to the Gemma 4 family — Gemma 4 12B.

This model is different from the Gemma 4 series released back in April (E2B, E4B, 26B-A4B, 31B). It's a 12B-parameter dense model, built around one core use case: running multimodal inference locally on a laptop with 16GB of VRAM.

UnslothAI has already released GGUF quantized versions, ready to use out of the box.

01. What does "encoder-free" mean? And why does it matter?

This is the most important concept to grasp from this post.

Traditional multimodal models process an image like this:

Image → Vision encoder (e.g., SigLIP/CLIP, 150M~550M parameters) → Extract feature vectors → Feed into the LLM

Audio works the same way: Audio → Audio encoder (300M parameters) → Extract features → Feed into the LLM

Gemma 4 12B skips the middleman entirely.

Image and audio data go straight into the LLM's main network, and the Transformer learns how to understand these non-text inputs on its own. No separate SigLIP vision tower, no standalone audio encoder — **the entire model is a unified Tr