Stability AI launches StableCode, a large language model for writing code

CTOnews.com, August 9, Stability AI is known for its Stable Diffusion text generation image model, but this is not the only area of interest, and the company is now dabbling in code generation. Today, Stability AI released its new Open large language Model (LLM) StableCode, which is designed to help users generate programming language code based on the transformer neural network.

StableCode will provide three different levels: a basic model for general purpose, an instruction model, and a long context window model that can support up to 16000 token.

The StableCode model benefits from the initial programming language datasets of the open source BigCode project, with additional filtering and fine-tuning by Stability AI. StableCode will first support the development of programming languages such as Python, Go, Java, JavaScript, C, markdown and C++.

Training any LLM depends on data, and StableCode's data comes from the BigCode project. CTOnews.com noted that HuggingFace and ServiceNow also launched an open BigCode-based StarCoder LLM in May. Nathan Cooper, chief research scientist at Stability AI, said StableCode's training involved extensive filtering and cleaning of BigCode data.

The long-context window version of StableCode has a 16000-word context window, and Stability AI claims to be larger than any other model, with a longer context window that allows for more professional and complex code generation prompts, and means that users can have StableCode view a medium-sized code base containing multiple files to help understand and generate new code.

StableCode is still in its early stages, and the official goal now is to see how developers will accept and use the model.