DeerFlow 2.0 is not an upgrade. It is a complete rewrite that shares no code with version 1. If you built on the original deep research framework, that codebase lives on the 1.x branch and still accepts contributions. Active development has moved entirely to 2.0.
The new architecture is built around a concept ByteDance calls a "super agent harness." The system orchestrates sub-agents, memory, and sandboxes to handle broad tasks, with extensibility driven by skills you can add or swap out. That design shift matters for builders: instead of a single research-focused pipeline, you now have a more general coordination layer you can point at different problem types.
On February 28, 2026, the day version 2 launched, DeerFlow reached the number one position on GitHub Trending. That kind of traction signals a real developer audience watching this project, which affects how seriously you should evaluate it as a dependency.
ByteDance recommends running DeerFlow with three models: Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5. That is a specific, concrete list worth noting if you are making model selection decisions today. The project is tied to ByteDance's Volcengine Coding Plan, and BytePlus has integrated its own intelligent search and crawling toolset called InfoQuest directly into DeerFlow. InfoQuest is available for free online experience, so you can test that integration without committing to anything.
The stack requires Python 3.12 or higher and Node.js 22 or higher. The license is MIT, which keeps the door open for commercial use and forking. The official website at deerflow.tech hosts real demos if you want to see the agent harness in action before pulling the repo.
The framework is multilingual in its documentation, with READMEs in Chinese, Japanese, French, and Russian alongside English. That suggests ByteDance is pushing for broad adoption outside its core market.
What to do now: If you are evaluating agent orchestration frameworks, clone the 2.0 branch, not the 1.x branch. Run the demos on deerflow.tech first to understand what the sub-agent coordination model looks like in practice. If your stack includes DeepSeek v3.2 or Kimi 2.5, you are already aligned with the recommended model set. And if you need intelligent search and crawling baked in, test the InfoQuest integration using the free tier before building your own tooling around it.