CodeShell-7B

CodeShell-7B is a multi-language code large language model developed by the Knowledge Computing Lab of Peking University under the WisdomShell organization. Trained from scratch on 500 billion tokens with a GPT-2-style architecture augmented by grouped-query attention and RoPE, it was positioned as an efficient code base model with strong HumanEval and MBPP results for its size.

What sets CodeShell apart is the full-stack ecosystem around it: the team also open-sourced IDE plugins for VS Code and JetBrains, fill-in-the-middle support for code completion, and a chat-tuned variant. Weights are distributed under Apache 2.0 terms, though users should confirm the exact license on the model card.