CodeGen is a family of open code generation models from Salesforce Research, described in “CodeGen: An Open Large Language Model for Code with Multi-Turn Program Synthesis” (arXiv, March 2022, by Erik Nijkamp and colleagues). The models scale up to 16.1 billion parameters and were trained on a mix of natural language and programming language data.
A central idea in the paper is multi-turn program synthesis: instead of asking the model to produce a whole program from one prompt, a user factors the task into a sequence of smaller prompts, each specifying a subproblem. The authors introduced an open benchmark called the Multi-Turn Programming Benchmark, with 115 problem sets designed for this step-by-step style, and found that spreading the same intent across several conversational turns significantly improved synthesis quality compared with a single prompt. Salesforce also released the JAXFORMER training library and the model checkpoints.
CodeGen was an early demonstration that conversational, incremental specification of a program could outperform one-shot generation, foreshadowing the interactive coding assistants that became common later. For business readers it highlighted that how you talk to a code model, breaking work into steps, can matter as much as the model’s raw size.