CodeGeeX

Type: full-code · Vendor: Z.ai (formerly Zhipu AI) / Tsinghua KEG · Language: Python (model), TypeScript (extensions) · License: Apache-2.0 (extensions + open-source models); proprietary hosted tier · Status: active · Status in practice: mature

Links: homepage docs repo

Open-source multilingual code generation extension from Tsinghua KEG / Zhipu, with completion, translation between languages, and an agent mode on top of the CodeGeeX4 (GLM-4-9B) model.

Description. CodeGeeX is an open-source coding assistant family launched in 2022 from Tsinghua University's Knowledge Engineering Group, now commercialised by Z.ai (formerly Zhipu AI). It ships VS Code, JetBrains, and Visual Studio extensions and exposes its models openly under Apache-2.0 (including CodeGeeX4-ALL-9B, a 128K-context coding model on GLM-4-9B that scored 82.3% on HumanEval). Modes include code completion, function-level generation, cross-language translation, code explanation, and an agent mode for multi-step coding tasks. Distinct from most coding assistants by virtue of being genuinely open-source end-to-end (model + extension + paper).

Agent loop shape. IDE extension that calls a hosted or local CodeGeeX model per user action. Completion is single-shot; agent mode runs a short ReAct loop over IDE tools (read, write, run). Repository indexing uses repository-level retrieval-augmented generation when context exceeds the 128K window.

Primary use cases

open-source self-hostable code completion (Apache-2.0 model and extension)
cross-language code translation across 300+ programming languages
free-tier coding assistant for individual developers and students
academic research on coding model evaluation (HumanEval-X benchmark ships with it)

flowchart TD USER[Developer] --> EXT[CodeGeeX extension<br/>VSCode / JetBrains / VS] EXT --> MODE{Mode} MODE -->|Completion| MODEL[CodeGeeX4-ALL-9B<br/>128K context] MODE -->|Chat| MODEL MODE -->|Agent| LOOP[ReAct loop] LOOP --> TOOLS[IDE tools<br/>read / write / run] TOOLS --> LOOP EXT -.context > 128K.-> RAG[Repo-level RAG] RAG --> MODEL MODEL --> EXT EVAL[(HumanEval-X<br/>shipped with project)] -.evaluates.-> MODEL

Key concepts

CodeGeeX4-ALL-9B (docs) — 128K-context coding model on GLM-4-9B, open-sourced under Apache-2.0.
HumanEval-X → eval-harness — Multilingual code-eval benchmark released alongside CodeGeeX (KDD 2023).
Open-source end-to-end stack — Model weights + extension + paper all under Apache-2.0, distinct from most coding assistants.
Repository-level RAG → naive-rag — Retrieval-augmented prompting kicks in when context exceeds 128K window.