RecurrentGemma (Griffin architecture)
RecurrentGemma is a pair of 2B and 9B open models that use the Griffin hybrid of gated linear recurrences and local attention, with a fixed-size state for efficient long sequences.
Matches similarly sized Gemma baselines despite fewer training tokens, while using less memory at inference. The underlying Griffin paper (arXiv 2402.19427, 2024-02-29) claimed Llama-2 parity with over 6x fewer tokens and beat Mamba with its Hawk RNN.
- Date
- Thursday, 11 April 2024
- Lab
- Google DeepMind
- Kind
- open-weights
- Access
- open weights (restricted license)
arXiv v1 2024-04-11 is the report date; the models were released with CodeGemma around 2024-04-09 per Google (not verified here).
Sources
This record was checked against its sources on 6 October 2026. How we check