GEMMA 4 TECHNICAL REPORT

Summary from the TLDR AI newsletter

Gemma 4 introduces a new generation of open-weight multimodal language models featuring dense and Mixture-of-Experts architectures, with parameters ranging from 2.3B to 31B. It enhances computational efficiency and reasoning via innovations like a thinking mode, long-context efficiency, and a unified encoder-free architecture. These models achieve performance comparable to larger models across various benchmarks and are available under an Apache 2.0 license for widespread adoption.