CodeGen is Salesforce AI Research's family of open-source code-generation models, built to convert plain English instructions into working code across multiple programming languages. It ships in several sizes, from 350 million parameters up to 16 billion, including both multilingual and Python-specific variants, and was competitive with early versions of OpenAI's Codex on benchmark coding tasks when it launched.
Across its releases, CodeGen 1.0 in early 2022, CodeGen 2.0 in early 2023, and CodeGen2.5 in mid-2023, Salesforce kept improving quality and reducing latency and cost, with CodeGen2.5 notably matching the performance of the larger 16B model at just 7 billion parameters, a meaningful efficiency gain. Internally, it powers Einstein for Developers, fine-tuned specifically for Salesforce's Apex language, which the company says saves developers roughly 90 minutes a day.
It's free and open source, with over 600,000 monthly downloads, used both as a research baseline and as a practical model for teams that want a smaller, self-hostable code model rather than relying on a much larger proprietary one.