Google has released CodeGemma, a family of open-source language models specialized for code generation and understanding. The models are designed to help developers write, complete, and debug code more efficiently.
CodeGemma comes in two sizes: a 2B parameter model for fast completions and a 7B parameter model for more complex tasks. Both are built on top of Google's Gemma architecture and are available on platforms like Hugging Face and Google's Vertex AI.
According to Google, CodeGemma was trained on a large corpus of code from public repositories, covering multiple programming languages including Python, JavaScript, Java, and C++. The models can handle tasks such as code completion, code generation from natural language descriptions, and code translation between languages.
Early benchmarks show that CodeGemma performs competitively with other open-source code LLMs like CodeLlama and StarCoder, particularly on the HumanEval and MBPP coding benchmarks.
"CodeGemma aims to make coding more accessible and productive for developers of all skill levels," said a Google spokesperson. "We're excited to see how the community will build upon these models."
The release includes model weights, training code, and example usage under a permissive license, allowing for both research and commercial use.