Google announces that it will add 'CodeGemma' with enhanced code completion, generation and command following, and 'RecurrentGemma' with optimized efficiency to its Gemma family of large-scale language models that are open and commercially available
Google has announced that it will add two new models, 'CodeGemma' and 'RecurrentGemma,' to the Gemma family of open, commercially available large-scale language models released in February.
Gemma Family Expands with Models Tailored for Developers and Researchers - Google for Developers
Gemma is a lightweight and open AI model created using the same technology as Gemini, an AI model also created by Google. In February, the base 'pre-trained model' and the 'instruction-tuned model' with improved dialogue performance were released.
Google releases Gemma, an open, commercially available, lightweight large-scale language model - GIGAZINE
The two new features announced this time are 'CodeGemma,' which enhances code completion, generation tasks, and command following, and 'RecurrentGemma,' which optimizes efficiency.
The CodeGemma model has one 2 billion parameter variant and two 7 billion parameter variants: the 2 billion parameter variant is suitable for fast code completion on the local computer, and the 7 billion parameter variant is specialized for code completion and code generation tasks, and one specialized for code-related chat and instruction following.
The results of comparing CodeGemma to other models are shown in the figure below. We compare its performance against competing models on single and multi-line code completion tasks.
RecurrentGemma is a memory-efficient model that leverages recurrent neural networks and local attention, achieving similar benchmark scores to the Gemma 2B model while reducing memory usage and significantly improving throughput.
Compared to the Gemma 2B model on the left, the RecurrentGemma 2B model on the right shows that throughput can be maintained even with longer sequences.
Like regular Gemma, CodeGemma and RecurrentGemma are open models and can be used commercially. Google has also released version 1.1 of the original Gemma model at the same time as releasing the two models.
Related Posts:
in Software, Posted by log1d_ts