StarCoder2 Unveiled: Open-Source LLM That Outperforms Its Predecessor with Fewer Parameters

ServiceNow, Hugging Face and NVIDIA have jointly released StarCoder2, a new open LLM series designed for developers.

Trained on 619 programming languages, StarCoder2 provides code generation, workflow creation and text summarization capabilities.

The model family includes three variants:

A 3 billion‑parameter model trained by ServiceNow.

A 7 billion‑parameter model trained by Hugging Face.

A 15 billion‑parameter model trained by NVIDIA.

Despite having only 3 billion parameters, the smallest model matches the performance of the original 15 billion‑parameter StarCoder, demonstrating that fewer resources can achieve comparable results.

Developers can fine‑tune these models with open‑source tools such as NVIDIA NeMo or Hugging Face TRL to meet specific needs.

“StarCoder2 proves the combined power of open‑science collaboration and responsible AI practices, delivering state‑of‑the‑art open‑access models that boost developer productivity and democratize access to code‑generation AI for organizations of any size,” said Harm de Vries, head of the StarCoder2 development team and co‑lead of BigCode.

“The joint effort led by Hugging Face, ServiceNow and NVIDIA enables a powerful foundation model that the community can use to build diverse applications with full data and training transparency,” added Leandro von Werra, machine‑learning engineer at Hugging Face and co‑lead of BigCode.

StarCoder2 is released under the BigCode Open RAIL‑M license, allowing free access and use for developers.