RecurrentGemma-2B
FreeOpen weights language model from Google DeepMind, based on Griffin
FreeFree tier
Inputs: textOutputs: text
About RecurrentGemma-2B
RecurrentGemma is a family of open-weights language models developed by Google DeepMind, based on the novel Griffin architecture. This architecture significantly improves inference speed for long sequences by replacing global attention with a mixture of local attention and linear recurrences. The repository provides model implementations in Flax (optimized) and PyTorch (reference), along with examples for sampling and fine-tuning. Pre-trained checkpoints (2B parameters) are available on Kaggle, and the project includes Colab tutorials for JAX and PyTorch sampling as well as JAX fine-tuning.
Key Features
Novel Griffin architecture with local attention and linear recurrences
Fast inference for long sequences
Open-weights model (2B parameters)
Optimized Flax implementation and reference PyTorch implementation
Examples for sampling and fine-tuning
Colab tutorials for JAX and PyTorch sampling, plus JAX fine-tuning
Available via Kaggle with model weights and tokenizer
Supports Poetry and pip installation
Pros & Cons
Pros
- Fast inference for long sequences due to Griffin architecture
- Open-weights allowing full access and modification
- Multiple framework support (Flax optimized, PyTorch reference)
- Comprehensive example scripts and Colab notebooks
- Backed by Google DeepMind with published technical report
Cons
- PyTorch implementation is un-optimized reference only
- Requires downloading model weights from Kaggle with a Kaggle account
- Limited to 2B parameter model size; no larger variants mentioned
- Dependency on JAX or PyTorch with specific installations
Best For
Language generation and samplingFine-tuning for custom tasks (e.g., English-to-French translation)Research into efficient recurrent architectures for language modelsDeploying fast inference for long-context applications
FAQ
What architecture does RecurrentGemma use?
RecurrentGemma is based on the Griffin architecture, which replaces global attention with a mixture of local attention and linear recurrences for faster inference on long sequences.