StableCode

StableCode, developed by Stability AI, is a generative AI large language model (LLM) designed specifically for coding assistance. It was first released as Stable Code Alpha 3B, an open-source model aimed at helping developers write code more efficiently. The model is trained on 560 billion tokens of code from the BigCode stack-dataset (v1.2) and further fine-tuned on popular languages such as Python, Go, Java, JavaScript, C, markdown, and C++. StableCode comes in three variants: a base model for general code generation, an instruction-tuned model for solving specific coding tasks using approximately 120,000 code instruction/response pairs, and a long-context-window model supporting up to 16,000 tokens, allowing users to work with the equivalent of five average-sized Python files simultaneously. The tool is intended to boost developer productivity by providing context-aware autocomplete suggestions and the ability to generate code from natural language instructions. Stability AI positions StableCode as a learning tool for new developers and a productivity enhancer for experienced programmers, aiming to make coding more accessible to a global audience. Performance comparisons using the HumanEval benchmark are provided to show its effectiveness relative to other models of similar size.