Google Releases CodeGemma AI Model for Real-Time Code Completion Speeding Up Programmer’s Code Editing

Google has released the LLM AI model for coding under the name CodeGemma, focusing on assisting in code writing. The model comes in three versions:

CodeGemma 2B: Specializing in code completion at 100%
CodeGemma 7B: Code completion at 80% and general information at 20%
CodeGemma 7B Instruct: Enhanced from the 7B model for command-based chat interaction

The training approach for CodeGemma utilizes the fill-in-the-middle (FIM) method. When measuring the ability to fill in code lines accurately, CodeGemma 2B performs at 78.41%, surpassing similar code-writing models while requiring significantly less runtime. On average, code completion takes less than 0.5 seconds (totaling 543 seconds from 1033 prompts) on a g2-standard-4 machine, making it practical to use.

While CodeGemma excels in accuracy, it also competes favorably with high-performance models like DeepSeek Coder in terms of speed.
Source: Google AI for Developers

TLDR: Google introduced the LLM AI model CodeGemma for coding, offering specialized code completion with efficient runtime and high accuracy compared to similar models like DeepSeek Coder.