DeerFlow 2.0 shares no code with its predecessor. ByteDance rebuilt the whole thing from scratch, and the scope has shifted significantly. Where version 1 was a Deep Research framework, version 2 is described as an open-source super agent harness that orchestrates sub-agents, memory, and sandboxes to do almost anything, powered by extensible skills.
The launch landed hard. On February 28th, 2026, DeerFlow claimed the number one spot on GitHub Trending the same day version 2 shipped. The original 1.x branch remains maintained for anyone still building on it, but active development has fully moved to 2.0.
The architecture change matters for builders. You are no longer wiring up a research pipeline. You are composing agents that can delegate to sub-agents, persist memory across tasks, and execute inside sandboxes. The extensible skills layer is what lets you expand the surface area without forking core logic.
ByteDance recommends running DeerFlow with Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5. Those are the three models called out directly in the source material, connected to their Volcengine Coding Plan offering.
On the tooling side, DeerFlow 2.0 now integrates InfoQuest, an intelligent search and crawling toolset independently developed by BytePlus. A free online experience is available for developers who want to try it before committing. This fills a real gap: agent workflows that need live web data have historically required custom glue code or third-party integrations. Having a first-party crawling layer baked in simplifies that considerably.
The stack requires Python 3.12 or higher and Node.js 22 or higher. Both are recent-enough requirements that you will want to check your environment before jumping in. The project is MIT licensed.
The rewrite also comes with a new official website at deerflow.tech, where the team has published real demos. That is worth checking before you read the README, since the demos give you a faster read on whether the 2.0 architecture fits your use case.
If you are building anything that involves chaining agents, delegating subtasks, or running code inside a controlled sandbox, this is worth a serious look today. Fork the repo, run through the demos on the site, and test against the recommended models. If you were already on DeerFlow 1.x, treat this as a migration decision, not an upgrade. The codebases are completely separate.