#Gemma

'Cache-to-Cache (C2C) lets LLMs exchange semantic information via KV-Cache fusion, boosting accuracy by about 3–10% over text-based pipelines and roughly halving latency.'

Google AI and DeepMind released VaultGemma 1B, the largest open-weight LLM trained end-to-end with differential privacy, showing that strong privacy guarantees can be combined with practical model performance