CodeFuse

Type: full-code · Vendor: Ant Group (蚂蚁集团) · Language: Python · License: open-source (MIT for components) · Status: active · Status in practice: mature

Links: homepage repo

Ant Group's open-source code LLM family covering the full software development lifecycle (design, requirements, coding, testing, deployment, operations) with both pre-trained models and downstream agent products.

Description. CodeFuse is Ant Group's umbrella project for Code LLMs and coding agents spanning the full SDLC. The codefuse-ai GitHub org hosts pre-trained code models (CodeFuse-13B, -CodeLlama-34B, CodeFuse-MFT-VLM for multimodal), evaluation harnesses (CodeFuseEval), and downstream agent components (Muagent for multi-agent orchestration, ModelCache for cache, ChatBot UI). Distinct from a pure code-assistant by treating the SDLC as a coordinated multi-stage workflow rather than a single completion loop.

Agent loop shape. Multi-component stack. The base layer is a coding LLM (CodeFuse-CodeLlama-34B, etc.). Muagent layers a multi-agent orchestration (planner, coder, reviewer, executor roles) on top, with ModelCache to avoid duplicate LLM calls. Each agent runs a tool-augmented ReAct loop scoped to its SDLC stage.

Primary use cases

open-source code LLMs for fine-tuning and research
multi-agent SDLC orchestration (Muagent) for full lifecycle automation
code-eval benchmark (CodeFuseEval) for in-house model evaluation
self-hosted enterprise coding agent for Ant Group's internal products

flowchart TD REQ[SDLC request] --> MUAGENT[Muagent orchestrator] MUAGENT --> PLANNER[Planner role] PLANNER --> CODER[Coder role] CODER --> REVIEWER[Reviewer role] REVIEWER --> EXEC[Executor role] EXEC --> ART[Code / tests / docs] REVIEWER -->|reject| CODER CODER --> MODEL[CodeFuse Code LLMs<br/>13B / 34B / VLM] CACHE[(ModelCache<br/>semantic cache)] -.dedup calls.-> MODEL EVAL[(CodeFuseEval)] -.evaluates.-> MODEL

Key concepts

Muagent → orchestrator-workers (docs) — CodeFuse's multi-agent framework for SDLC orchestration.
CodeFuseEval → eval-harness — Code-eval harness shipped alongside the models.
ModelCache → prompt-caching — Semantic cache layer for LLM calls to reduce duplicate inference.