StarCoder2 15B Instruct

StarCoder2 15B Instruct is a 15-billion-parameter code generation model from the BigCode collaboration — HuggingFace and ServiceNow — released in September 2024 and trained on The Stack v2, a dataset restricted to permissively licensed source code. The 16K context window handles most single-file and small multi-file completion tasks cleanly. On HumanEval, it has been outpaced by DeepSeek Coder V2 and Qwen 2.5 Coder, so it is not the highest raw-throughput choice for benchmark-chasing. Its differentiated value is training-data provenance: teams in regulated industries or with strict IP policies around model training data favor StarCoder2 precisely because every training example carries a verified open-source license. Released under BigCode OpenRAIL-M, which restricts a narrow set of harmful applications but is otherwise commercially usable.