Google Releases Gemma 4 Models with Quantization-Aware Training

Google has released new checkpoints for its Gemma 4 family of models, utilizing Quantization-Aware Training (QAT). This method trains the models to be more accurate when their weights are compressed to very low bit-widths, such as 4-bit or even 2-bit for specific layers. The goal is to enable these models to run efficiently on consumer hardware with significantly reduced memory footprints, like the E2B model requiring only about 1 GB. AI

IMPACT Enables efficient on-device AI by significantly reducing model size and memory requirements.

RANK_REASON Frontier-lab model release with system card. [lever_c_demoted from frontier_release: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• Gemma 4
• GGUF
• Google
• llama.cpp
• Quantization-Aware Training

• model release
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →