Gemma 4 (2 articles)

Google Open Sources Gemma 4: KV Cache Compressed to 3 Bits, Saving 6 Times Memory; Comprehensive Performance Yet to Be Verified by Third Parties

Google recently launched the open-source multimodal AI model Gemma 4, introducing video and image processing capabilities. The new TurboQuant technology compresses KV caches to 3 bits, achieving significant memory savings.

Google's Open Source Gemma 4 Shocker: How a 31B Multimodal Model Challenges AI Giants' Dominance on Raspberry Pi?

In a groundbreaking move in 2026, Google DeepMind released the open-source Gemma 4 series, touted as the smartest open-source model family to date, designed for advanced reasoning and agent workflows. This release marks a significant leap in open-source large models toward multimodal and edge computing, although its "unconfirmed" status urges caution.