Hugging Face and ServiceNow drop a free Copilot rival

StarCoder is a 15-billion-parameter open-source model designed to generate code. In plain terms, it gives you a royalty-free, self-hostable alternative to GitHub Copilot that you can inspect and fine-tune yourself.

What it does

StarCoder supports over 80 programming languages and plugs directly into editors like Visual Studio Code. Beyond filling in code snippets on the fly, it can follow basic natural language prompts like "build an app UI" and answer questions about existing codebases.

According to co-lead Leandro von Werra, StarCoder matches or outperforms the initial OpenAI model that originally powered Copilot. While it does not launch with the full bells and whistles of commercial suites, it gives you full control over your development environment without sending proprietary code to outside servers.

How it was built

The model was developed under the BigCode initiative run by Hugging Face and ServiceNow Research, trained on an in-house cluster of 512 Nvidia V100 GPUs over several days.

Its training foundation is The Stack, a dataset containing more than 19 million permissively licensed repositories and 6 terabytes of code across 350 languages. Unlike proprietary rivals that scrape public code without clear guardrails, BigCode actively scrubbed the data.

The team stripped out personally identifiable information such as names, emails, IP addresses, and private API keys. They also used automated tools to discard files containing known malware or security exploits, while offering an opt-out system for developers who want their repositories excluded.

Licensing and availability

Everything is available on GitHub today, including the code repositories, training frameworks, filtering methods, and evaluation suites.

StarCoder is distributed under an OpenRAIL-M license, meaning anyone can use it commercially without paying royalties. The catch is legal and behavioral: derivatives are legally barred from generating malicious software or exploits. The creators also caution that the model can still occasionally generate inaccurate code or false logic, so human code review remains mandatory.

Why it matters

If your organization has banned GitHub Copilot over privacy fears or strict IP policies, StarCoder provides a legitimate way out. You can deploy it locally inside a private cluster, fine-tune it on internal repos, and keep sensitive engineering work completely behind your firewall.

For the broader ecosystem, it breaks open a segment that was quickly becoming a closed monopoly. By providing a clean, legally vetted base model, BigCode gives independent developers and startups the foundation to build specialized coding assistants without paying rent to OpenAI or Microsoft.