Google Unveils Code-Specialized Model 'CodeGemma'

Google has released the CodeGemma model family, specialized for coding tasks based on Gemma. These models were trained on approximately 500 billion additional tokens (English, math, and code data) to strengthen logical and mathematical reasoning capabilities.

CodeGemma is available in three versions:

• CodeGemma 2B: Specialized for code infilling, suitable for fast code completion and generation.
• CodeGemma 7B: Optimized for code understanding and generation through parallel training on code infilling and natural language.
• CodeGemma 7B Instruct: An instruction-following model for code-related conversations and mathematical reasoning.

All models support an 8K token context window. In terms of performance, CodeGemma 7B showed top-tier performance among similarly sized models on the HumanEval and MultiPL-E benchmarks, and recorded particularly outstanding results on the GSM8K benchmark.

Through integration with the Hugging Face ecosystem, it is immediately available on the Transformers library, Google Cloud, and Inference Endpoints.