Google Unveils CodeGemma: New AI Models for Code Generation & Reasoning

Google has launched an extended Gemma artificial‑intelligence model series, releasing two new variants: one for code generation and another for inference.

The code‑generation variant, called CodeGemma, provides intelligent code completion and generation, capable of producing entire code blocks in a single step.

According to Google, CodeGemma was trained on up to 5 trillion tokens from online documentation, mathematics, and code, and can work with many popular programming languages.

It comes in several versions, including a 7B pre‑trained model for code generation and completion, a 7B instruction‑tuned model excelling at code chat and instruction following, and a lightweight 2B pre‑trained model for fast local code completion.

CodeGemma is part of the open‑access Gemma family, fine‑tuned for software development scenarios.

The suite includes three distinct models:

A 2‑billion‑parameter base model optimized for low latency and privacy‑focused code completion.

A 7‑billion‑parameter base model that combines code completion with natural‑language processing for enhanced utility.

A 7‑billion‑parameter instruction‑following model designed to assist developers with code, programming, and mathematical reasoning conversations.

These models leverage pre‑trained Gemma checkpoints and an additional 5 trillion tokens covering English, mathematics, and various coding languages, setting new standards in logical and mathematical reasoning for code generation.

The 7‑billion‑parameter model performs strongly across languages such as Python, Java, JavaScript, and C++, and achieves top results on HumanEval, MultiPL‑E, and GSM8K benchmarks.

RecurrentGemma, aimed at improving inference with larger batch sizes, offers lower memory requirements, enabling higher token throughput on memory‑constrained devices.

Both models are currently available for trial on Kaggle, Hugging Face, and Vertex AI Model Garden.