Stability AI releases Stable Code 3B for enhanced coding assistance

18.01.2024

Stability AI has recently unveiled Stable Code 3B, an upgraded AI system with three billion parameters, designed to facilitate automatic code generation and completion in software development.

With notable improvements such as a larger context size and enhanced code completion quality, Stable Code 3B aims to redefine the capabilities of AI-assisted software development. The key distinction lies in its efficiency, as this model runs seamlessly on standard hardware like laptops, eliminating the need for specialized and costly chips. Despite its compact size, Stability AI asserts that Stable Code 3B not only matches but surpasses the code completion quality of models with more than double its parameters.

This advanced AI system builds upon Stability AI’s Stable LM natural language model, incorporating additional training data from software engineering sources such as code repositories and programmer forums. It boasts proficiency in 18 programming languages, including Python, JavaScript, Java, C++, and Go.

The model’s training process underwent optimization through the integration of Rotary Position Embeddings (RoPE), enhancing performance by expanding the context size. This technique, also utilized by Meta’s Llama 2 Long, allows for context lengths of up to 100k tokens.

Beyond merely suggesting new lines of code, Stable Code 3B introduces an advanced capability known as Fill in the Middle (FIM). This feature enables the system to automatically generate entire functions or components, addressing large missing sections in existing code.

The realm of AI-generated code has captured the attention of major tech players like Microsoft, OpenAI, and Meta. Stability AI’s latest release positions itself as a leader in this dynamic field, outperforming comparable models like StarCoder.