CodeGen is Salesforce’s family of open-source large language models for program synthesis โ AI models that turn natural-language intent into code. Trained on TPU-v4 pods, the CodeGen models were competitive with OpenAI’s Codex at their release and were an early proof that open code models could reach commercial quality. The Apache-2.0 repository has more than 5,100 stars on GitHub and saw renewed activity in 2026.
The CodeGen family comes in several flavours: CodeGen-NL, pretrained on natural language and code; CodeGen-Multi, trained across many programming languages; and CodeGen-Mono, further specialised on Python. Sizes ranged from 350M to 16B parameters, and Salesforce’s research emphasised multi-turn program synthesis โ generating code step by step through a conversational sequence of prompts rather than in one shot, which measurably improved results on their benchmarks. The repository provides the model code, checkpoints and evaluation setup, and the models are also on Hugging Face for use with standard tooling. Historically, CodeGen mattered beyond its own scores: released openly in 2022, it helped kick off the wave of open code LLMs that followed, and it remains a clean reference for how large code models of that era were trained and evaluated. Its conversational, multi-step view of synthesis also anticipated the agent-style coding workflows that are now mainstream.
Free. CodeGen is open source under the Apache-2.0 license (models carry the BSD-3-Clause terms noted in the repository).
Researchers and developers interested in code generation models โ as a usable open code LLM family and as a reference implementation from the early open-code-model era.