StarCoder is an open-source 15.5B-parameter code LLM (StarCoder / StarCoderBase) trained primarily on The Stack (v1.2) with an opt-out mechanism. It was trained with Fill-in-the-Middle (FIM) objective and uses multi-query attention to support long-context code tasks (commonly described as ≈8k context). StarCoderBase was trained on roughly 1 trillion tokens (paper) and StarCoder is a fine-tuned variant (StarCoderBase → StarCoder) with additional fine-tuning on ~35B Python tokens to improve Python performance. The project uses the Megatron-LM framework and standard mixed-precision training stacks (Megatron-LM, PyTorch, NVIDIA tooling/apex) with training and inference optimizations discussed on the model card. The release includes data-preprocessing code, an attribution tool, a PII-redaction pipeline, evaluation harness, and governance materials. The model is released under the BigCode OpenRAIL-M v1 license; certain weights/files are gated behind a license acceptance on Hugging Face. Typical uses include code completion, infilling, program editing, natural-language explanations of code, and building technical-assistant flows (e.g., via a Tech Assistant prompt). Benchmarks reported on the model card/paper include HumanEval, MBPP, DS-1000, and MultiPL-HumanEval (paper reports, for example, ~40% pass@1 on HumanEval under the reported prompting conditions). The model card and paper provide evaluation details, limitations, intended uses, and guidance for attribution and governance.