Nvidia, Hugging Face and ServiceNow release new StarCoder2 LLMs for code generation

Nvidia, Hugging Face, and ServiceNow have launched StarCoder2, a new series of open-access large language models (LLMs) designed to enhance AI for code generation. These models, available in three sizes, have been trained on over 600 programming languages to assist enterprises in accelerating various code-related tasks in their workflows. Developed as part of the BigCode Project, a collaborative effort between ServiceNow and Hugging Face, the models are offered royalty-free under Open Responsible AI Licenses (OpenRAIL).

Leading the development of StarCoder2, Harm de Vries highlighted the importance of open collaboration and ethical AI practices in enhancing developer productivity and enabling organizations of all sizes to leverage the benefits of code generation AI. Unlike its predecessor, StarCoder2 features models in three different sizes – 3B, 7B, and 15B – trained on a wide array of programming languages, including low-resource ones like COBOL and mathematics.

Utilizing novel training techniques, the latest generation of models aims to understand and generate diverse programming languages efficiently. The 3B and 7B models leverage ServiceNow’s Fast LLM and Hugging Face’s nanotron frameworks, respectively, while the 15B model is optimized with Nvidia’s NeMo cloud-native framework and TensorRT-LLM software.

Although the performance of these models across coding scenarios is yet to be fully assessed, early tests show promising results. Notably, the 3B model’s performance matches that of the original 15B StarCoder LLM. Enterprises can choose and further fine-tune these models for specialized tasks such as code and workflow generation, text summarization, code completion, and more, enhancing their development processes.

With a focus on providing accurate and context-aware predictions, the StarCoder2 models enable accelerated development by saving time for engineers and developers to concentrate on critical tasks. These models are available under the Open RAIL-M license, encouraging broad access and responsible AI usage. Interested parties can access the models from Hugging Face or experiment directly with the 15B model on the Nvidia AI Foundation platform.

While StarCoder2 expands the options for AI-driven code generation, it faces competition from other industry players like OpenAI and Amazon, each offering their own AI coding tools. The growing adoption of LLMs in application development underscores the potential of AI technologies in enhancing efficiency and innovation across industries.