Gemma is a family of lightweight open models built from the research behind Gemini. The first release has 2 billion and 7 billion parameter models, each with pretrained and fine-tuned checkpoints, and the report covers training and safety evaluations.
Findings
- Better than similarly sized open models on 11 of 18 text tasks.
Limits
- Abstract-level reading only. The report evaluates text tasks only; the vision-language Gemma models used in later work are separate releases.
Relevance
Open weights at this size are what let a robot or an assistive agent run its reasoner on board, so camera and microphone data stay local, as in GoGet, which uses a later Gemma. Small local models also fail in their own ways, which is one reason to cross-check them against a second model. See foundation-models-for-robotics.
Links
- arXiv
- Subfield: Robot learning, Perception