Google DeepMind Releases RecurrentGemma: One of the Strongest 2B-Parameter Open Language Models Designed for Fast Inference on Long Qequences

Google DeepMind has released RecurrentGemma, a language model that achieves high performance and faster processing speeds for long text sequences while reducing memory usage. The model addresses the challenges in language model development by combining linear recurrences with local attention mechanisms. RecurrentGemma is capable of processing up to 40,000 tokens per second and is ideal for applications where resources are limited.