StarCoder gives developers and companies access to a code-generating AI model they can use without royalty payments. Built by Hugging Face and ServiceNow Research, it is intended to be adapted by the community, while its training choices and license reflect the risks that come with generating software from learned examples.
A model designed for wider access
StarCoder can follow basic instructions, answer questions about code and work with Microsoft’s Visual Studio Code editor. It was trained on over 80 programming languages, with material from GitHub repositories that included documentation and programming notebooks.
The project’s public release includes more than the model itself. The BigCode team made its code repositories, training framework, dataset filtering methods, evaluation suite and research notebooks available on GitHub. Researchers and developers can examine how the system was built, and the project expects community input to help improve it.
That access matters in a field where many code-generation tools are offered through commercial products or paid APIs. The article describes StarCoder as royalty-free for anyone, including corporations. Its co-lead Leandro von Werra said that developers could fine-tune and adapt the model for their own use cases, opening the door to applications beyond the initial release.
Training data and safeguards
StarCoder is a 15-billion-parameter model trained on The Stack, an open-source dataset containing over 19 million curated, permissively licensed repositories and more than six terabytes of code in over 350 programming languages. The article explains that model parameters are learned from training data and help define how the system performs tasks such as generating code.
The Stack’s permissive license allows code to be copied, modified and redistributed. BigCode also created a way for developers to opt out of the dataset. The team worked to remove personally identifiable information, including names, usernames, email and IP addresses, keys and passwords. It created a separate collection of 12,000 files containing personal information, which it plans to make available to researchers through gated access.
The team also used a Hugging Face tool to detect malicious code and remove files that might be considered unsafe, including files with known exploits. These steps address risks in training data, but they cannot guarantee every generated answer will be safe. The release’s technical paper says the model may still produce inaccurate, offensive or misleading content, personal information, or malicious code that passed through filtering.
Use rights come with conditions
StarCoder’s license permits commercial use, but the model is not open source in the strictest sense. It is released under OpenRAIL-M, a license with legally enforceable use-case restrictions that also apply to derivatives and apps using the model. Users must agree not to use it to generate or distribute malicious code.
The restrictions establish expectations for users, but the article notes that technical safeguards cannot stop someone from ignoring the terms. That leaves an open question about how the license will be respected in practice. For teams considering the model, the distinction matters: free commercial use does not mean every use is allowed.
There are also questions around code ownership and deployment. Critics have raised concerns about AI systems trained on public code, especially when licenses and attribution may not be clear in generated output. The article notes that legal experts have warned that incorporating copyrighted or sensitive text into production software could create risks for companies.
Other tools have taken steps to address related concerns. GitHub added a setting to block suggestions that match public content from GitHub, while Amazon’s CodeWhisperer can show and optionally filter the license associated with similar suggested functions. StarCoder’s release makes its training and filtering work available for inspection, but developers still need to consider what the model produces and the conditions governing its use.
An early platform, not a finished product
StarCoder’s developers say the initial release will have fewer features than GitHub Copilot. The project’s case for releasing it anyway is that researchers and the wider community can study, adapt and build on a capable system without needing the resources or expertise to train one themselves.
ServiceNow Research has described the project as a way to build expertise in responsible generative AI and provide transparency to developers and customers. The article says the company did not disclose its financial investment, though it described its donated computing resources as substantial. How StarCoder may eventually fit into ServiceNow’s products remains open.
For developers, the release offers both an adaptable coding tool and a view into how it was assembled. Its license, dataset controls and safety filtering provide a framework for responsible use, while the possibility of inaccurate, sensitive or malicious output remains. The practical test will be how well developers evaluate its suggestions and follow the license conditions as the community develops new applications.