Stability AI has announced the release of Stable Code 3B, an upgraded three billion parameter AI system for automatic code generation and completion.

With enhancements like larger context size and improved completion quality, Stable Code 3B aims to push the boundaries of AI-assisted software development.

At just three billion parameters, Stable Code 3B is designed to run efficiently on readily available hardware like laptops—unlike larger models which require expensive specialised chips. Despite its smaller size, the company claims it matches or exceeds the code completion quality of models over twice its size. 

The system builds on Stability AI’s Stable LM natural language model with additional training on software engineering data like code repositories and programmer forums. It covers 18 programming languages including Python, JavaScript, Java, C++, and Go.

The model’s training process witnessed optimisation through the incorporation of Rotary Position Embeddings (RoPE), expanding the context size for improved performance. This technique, also employed by Meta’s Llama 2 Long, allows for context lengths up to 100k tokens.

Beyond simply suggesting new lines of code, it can also fill in large missing sections in existing code. This advanced abilit