Qwen Code 1.0: Alibaba Open-Sources Its Agent Harness

Alibaba released Qwen Code 1.0 on November 11, folding Qoder's agent harness into the CLI and open-sourcing the result under Apache 2.0. The headline is not the version number. It is that a terminal coding assistant has turned into an agent runtime you can embed in your own software — and that the free tier people remember is not coming back.

What actually shipped

Per the release write-up, "Alibaba merged Qoder's agent harness capabilities into Qwen Code and open-sourced the whole thing together as Qwen Code 1.0." Qoder was Alibaba's agent-assisted development tool; the harness is the machinery that plans, calls tools, and keeps an agent on track across a long task.

The architecture, as described in Qwen Code's documentation and relayed by the write-up, "starts from an in-process, in-memory event bus and evolves toward a distributed multi-node, multi-session design, with a unified event model that decouples reasoning orchestration from tool execution." Read plainly: planning and tool execution are separate, and the events between them are observable rather than buried in a loop.

The caveat in the docs

The official docs carry a "Qwen Code 1.0 Preview" note. The 1.0 Preview is a separate branch and the stable branch is retained — so this is a direction of travel, not a finished release. Teams that need stability should keep the stable branch installed alongside it.

Installing it, and what it costs

Install is unchanged: npm install -g @qwen-code/qwen-code, or the standalone shell script on Linux and macOS. First launch asks you to connect a model provider, and the list is broad — Qwen, OpenAI-compatible providers, Anthropic, Gemini, and local providers. A VS Code extension is in beta for people who would rather not live in the terminal.

The money question has a blunt answer. Model usage is billed by whichever provider you connect, and the free Qwen OAuth tier ended in April 2026. A Qwen Plus or Pro subscription unlocks Qwen Code and Agent access; otherwise you pay per token through QwenCloud, ModelStudio, or your own provider. The software is free and open source. Running it is not free.

Should you switch?

If you are…	What this release changes
A daily CLI user on Qwen	Install the preview branch, keep the stable one. The everyday fix-this-bug workflow is unchanged.
Building your own agent product	This is the reason to look: a production-used harness, open source, with an event model you can hook into.
Happy on Claude Code or Cursor	Nothing here forces a switch. The write-up puts it bluntly: the value "is for builders, not for people who just want a better autocomplete."
Waiting for the free tier	It did not return. Budget for tokens.

The sceptical read is worth keeping: an embeddable runtime is only as good as its docs, stability guarantees, and community, and a preview branch has none of those settled. Apache 2.0 licensing lowers the legal risk of betting on it; it does nothing about breaking changes.

What to watch

Alibaba shipped two things the same day — Qwen Code 1.0 and the Qwen4-Max preview — which reads as a platform play rather than a tooling update: own the runtime agents are built on, not just the model underneath someone else's. The test is whether third-party products actually embed the harness and survive its preview-stage churn.

It lands in a week when the agent-harness layer is moving generally. Cursor's Projects feature went multiplayer in beta for all users on November 10, putting a shared context layer over each engineer's agents — though the launch write-up notes that Cursor "publishes no numbers on how often coordinators actually catch the overlap." Two different bets on the same premise: the harness around the model is where the next round of differentiation happens. Neither has shipped evidence yet.