Google DeepMind has released Gemma 4, a new series of open-source AI models featuring four vision-capable reasoning models under Apache 2.0 license. The models use Per-Layer Embeddings (PLE) technology to maximize parameter efficiency, making them particularly suitable for on-device deployments. The release includes 2B, 4B, 31B, and a 26B-A4B Mixture-of-Experts model, with Google emphasizing unprecedented intelligence-per-parameter performance.
Background
Open-source AI models have been rapidly evolving, with significant focus on creating smaller, more efficient models that can run on consumer hardware while maintaining strong performance capabilities.
- Source
- Simon Willison
- Published
- Apr 3, 2026 at 02:28 AM
- Score
- 7.0 / 10