gemma-4-12b-qat

Gemma 4 12B QAT

Gemma 4 12B QAT is the Quantization-Aware Training version of Gemma 4 12B. It aims to keep quality close to bfloat16 while using much less memory to load the model.

Gemma 4 is an open multimodal model family from Google DeepMind. It supports text and image input, text output, reasoning, long context, system prompts, and native tool use.

Gemma 4 12B uses the Unified design, which routes multimodal inputs into the decoder-only LLM backbone through lightweight projection layers instead of separate encoders. The QAT build keeps that architecture while reducing the memory needed to load the model.