CodeGen

Why we included this project

The CodeGen models from Salesforce AI Research span three generations, from the original CodeGen1 through CodeGen2 and CodeGen2.5, with checkpoints ranging from 350M up to 16B parameters. They turn natural-language descriptions into working code, which makes them a practical way to experiment with program synthesis locally instead of depending on a closed API. CodeGen2 brought fill-in-the-middle infill generation, and the 7B CodeGen2.5 checkpoint is worth a look because Salesforce reports it outperforms 16B models, giving you a sensible spread of sizes to match whatever GPU you have. Every checkpoint loads directly through the Hugging Face transformers API in a few lines of Python, so self-hosting is straightforward. Just remember this is research software rather than a supported product, so budget time to test the output on your own workload before trusting it in production.