CodeFuse
Ant Group's open-source code intelligence stack (蚂蚁 CodeFuse): an Apache 2.0 GödelScript static analysis engine, MIT-licensed Ling-Coder-Lite weights and an OpenSumi-based AI IDE.
Overview
Start with the caveat, because it decides whether this deserves your time. Ant Group's CodeFuse is a research-led open-source stack rather than a product with a release train. Papers keep landing, two on the CodeFuse toolchain at FSE 2026, two more at ICML 2026 and another at KDD 2026, while the shipped artifacts have gone quiet. CodeFuse IDE stopped at 0.7.0. CodeFuse-Query's last tagged release is 2.1.0. Ling-Coder-Lite has not been refreshed since 27 March 2025. Commits continue across the organisation through August 2026, so this is slow rather than dead, but plan around that.
The component that justifies a look is CodeFuse-Query, licensed Apache 2.0, which treats static analysis as a database problem. Source compiles into relational facts and you query it with GödelScript, a Datalog dialect, rather than writing yet another AST visitor. Eleven languages are covered, with Java, JavaScript, TypeScript, XML and Go at mature status and Objective-C, C++, Python3, Swift, SQL and Properties in beta. For anyone building code-quality or security tooling across a large polyglot monorepo this is a meaningfully different approach from what the Western ecosystem offers, and it was built against Ant's own codebase rather than as a demo.
On the model side, Ling-Coder-Lite is MIT licensed at 16.8 billion total parameters with 2.75 billion activated, published to both Hugging Face and ModelScope. Plain MIT is unusually clean for a Chinese lab release, with no user cap and no acceptable-use annex, which is exactly the kind of detail this catalogue exists to get right. The limitation is the 16K context window, fine in March 2025 and not enough for agentic work now. CodeFuse-CGM, the graph-based framework accepted at NeurIPS 2025, is the more interesting research artifact for anyone working on repository-level reasoning. Practical warning: documentation is mostly Chinese, the English READMEs are minimal and there is no single entry point across the repositories, so budget reading time before committing.
Key Features
- ✓ CodeFuse-Query, Apache 2.0, a static analysis engine that compiles source into relational facts and queries them with GödelScript, a Datalog dialect, instead of hand-written AST visitors
- ✓ Eleven languages covered, with Java, JavaScript, TypeScript, XML and Go at mature status and Objective-C, C++, Python3, Swift, SQL and Properties in beta
- ✓ Ling-Coder-Lite, MIT licensed, a mixture-of-experts code model at 16.8B total parameters with 2.75B activated and a 16K context window, released 27 March 2025 to Hugging Face and ModelScope
- ✓ CodeFuse-CGM, a graph-based LLM framework for repository-level software engineering tasks, accepted at NeurIPS 2025
- ✓ MFTCoder multi-task fine-tuning framework (KDD 2024) and muAgent, a knowledge-graph-driven agent framework
- ✓ CodeFuse IDE, an AI-native editor built on Ant's OpenSumi framework with VS Code extension compatibility, shipping Windows x64, macOS x64 and macOS arm64 builds
Where it holds
- • Licensing is unambiguous and generous: Apache 2.0 across the tooling and plain MIT on the Ling-Coder-Lite weights, with no acceptable-use annex and no commercial application process
- • CodeFuse-Query takes a genuinely different approach to static analysis, treating a codebase as a queryable relational database rather than a parse tree to walk
- • Backed by real production scale at Ant Group and by a continuing research output, with papers accepted at FSE 2026, ICML 2026 and KDD 2026
- • Weights published to Hugging Face and ModelScope together, so mainland teams are not blocked on a mirror
Where it breaks
- • Release cadence has slowed sharply. CodeFuse IDE has not tagged past 0.7.0, CodeFuse-Query sits at 2.1.0 and Ling-Coder-Lite has not been refreshed since 27 March 2025, though commits continue across the organisation through August 2026
- • Ling-Coder-Lite's 16K context window was adequate in early 2025 and falls well short of what agentic coding work needs in 2026
- • Documentation is predominantly Chinese. English READMEs are thin, and GödelScript carries a real learning curve with little English teaching material
- • The work is fragmented across more than a dozen repositories with no single entry point, so evaluating what is current takes meaningful reading time
My Take
Release cadence is the problem here. CodeFuse IDE has not tagged past 0.7.0, CodeFuse-Query sits at 2.1.0 and Ling-Coder-Lite has not moved since 27 March 2025, though 2026 commits plus accepted papers at FSE and ICML show the team is still working. What survives that is CodeFuse-Query, an Apache 2.0 engine that compiles source into relational facts and queries them in GödelScript, a Datalog dialect, across eleven languages. Worth a weekend if you build code-quality tooling for a large polyglot monorepo, not worth betting a roadmap on.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0 (CodeFuse-Query and most repositories), MIT (Ling-Coder-Lite weights)
- Starting at:
- Free. Every component is published on GitHub under Apache 2.0 or MIT with no commercial gate, no user cap and no separate model licence to apply for. Ling-Coder-Lite weights are mirrored on Hugging Face and ModelScope. There is no hosted paid tier, so the cost is entirely your own inference hardware and engineering time.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where CodeFuse holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to CodeFuse
Qwen3-Coder
open-sourceAlibaba's open-weights coding model family plus the Apache 2.0 Qwen Code terminal agent, separate artifacts from the hosted Qwen Chat product. Qwen3-Coder-Next runs 80B total, 3B active.
Ling 3.0
open-sourceAnt Group's open-weight MoE family (百灵, Bailing): Ling-3.0-flash at 124B total and 5.1B active, a 7.9B tiny variant and trillion-parameter Ring reasoning models, all MIT.
Semgrep
freemiumStatic analysis and AppSec platform where rules are readable code patterns, plus a plugin that scans every file an AI coding agent writes.