Gemma 4 12B: The Developer Guide

Gemma 4 12B introduces an encoder-free multimodal architecture that directly processes vision, audio, and text through a single transformer backbone.

• •Vision embedder (35M parameters) replaces 27-layer vision transformers by projecting raw 48x48 pixel patches directly to LLM hidden dimensions
• •Audio wave projection eliminates separate audio encoders by converting 16 kHz signals into 40ms frames for direct LLM input
• •Unified fine-tuning allows downstream adapter tuning (LoRA) or full model tuning in a single pass without freezing separate encoders
• •Achieves automatic speech recognition, agentic reasoning, video understanding, and coding with 16GB VRAM local inference capability
• •Provides OpenAI-compatible local API servers via LiteRT-LM CLI for drop-in integration with existing developer tools