Nvidia, Hugging Face, and ServiceNow have launched StarCoder2, an advanced suite of open-access large language models (LLMs) designed to enhance AI-driven code generation.
Currently available in three sizes, these models have been trained on over 600 programming languages, including lesser-known ones, enabling enterprises to streamline various coding tasks within their development workflows. Developed as part of the open BigCode Project—a collaborative initiative by ServiceNow and Hugging Face—the models promote responsible use of LLMs for code and are provided royalty-free under Open Responsible AI Licenses (OpenRAIL).
“StarCoder2 exemplifies the strength of open scientific collaboration combined with responsible AI practices,” stated Harm de Vries, lead of ServiceNow’s StarCoder2 development team. “This state-of-the-art model boosts developer productivity and democratizes access to code generation AI, allowing organizations of all sizes to realize their full business potential.”
StarCoder2: Three Models for Diverse Needs
The original StarCoder LLM featured a single 15B-parameter model trained on 80 programming languages. In contrast, StarCoder2 introduces three distinct sizes—3B, 7B, and 15B—trained on 619 programming languages. The training dataset, known as The Stack, is over seven times larger than its predecessor.
Significantly, new training techniques have been implemented to enhance the models’ ability to comprehend and generate code in low-resource languages, such as COBOL, as well as mathematical expressions and program source code discussions.
The 3B model utilizes ServiceNow’s Fast LLM framework, while the 7B model employs Hugging Face’s nanotron framework. Both are designed for high-performance text-to-code and text-to-workflow generation while minimizing computing demands. Meanwhile, the 15B model is optimized using the Nvidia NeMo cloud-native framework and Nvidia TensorRT-LLM software.
Collaborative Innovation: ServiceNow, Hugging Face, and Nvidia
While the performance of these models across various coding scenarios remains to be seen, initial tests suggest that the 3B model performs comparably to the original 15B StarCoder LLM. Enterprise teams can customize any of these models using their organizational data for specific applications, including source code generation, workflow automation, code completion, advanced summarization, and snippet retrieval.
The models' extensive training enhances their ability to provide accurate and context-aware predictions, thus accelerating development processes and allowing engineers to focus on more critical challenges.
“Every software ecosystem features a unique programming language, and code LLMs can foster significant advancements in efficiency and innovation across industries,” noted Jonathan Cohen, vice president of applied research at Nvidia. “Our partnership with ServiceNow and Hugging Face delivers secure, responsibly developed models that broaden access to accountable generative AI for the global community.”
Getting Started with StarCoder2
All models in the StarCoder2 family are available under the Open RAIL-M license, providing royalty-free access. Supporting code can be found on the BigCode project’s GitHub repository, and the models are also downloadable from Hugging Face. Additionally, the 15B model is accessible via Nvidia AI Foundation, enabling developers to experiment directly from their browsers or through an API.
While StarCoder represents a significant advancement in AI-driven code generation, it is not alone in the field. Competitors like OpenAI with Codex (which powers GitHub Copilot), Amazon’s CodeWhisper, and others such as Replit and Codenium are also exploring the capabilities of LLMs in application development.