Qwen Code 1.0 is a preview, not a finished upgrade. Alibaba's Qwen team published the v1.0.0-preview release of its open source terminal coding agent on Nov 11, 2026. It merges the agent harness from Qoder, Alibaba's commercial coding product, into Qwen Code, ships as a single native program that needs no Node.js, and adds Claude Code style permission modes. The stable 0.27 line keeps going on a separate branch. Try 1.0 in its own config folder if you live in the terminal; stay on 0.27 if you rely on the IDE plugins, the web shell, chat channels or a working CI script.

What Qwen Code 1.0 Preview actually is

Qwen Code started life as a fork of Google's Gemini CLI (version 0.8.2) and has developed on its own since its 0.1 release, under the Apache 2.0 license. The 1.0 Preview is a rebuild of its core. According to the v1.0.0-preview release notes, it:

  • brings the agent capabilities of Qoder and Qwen Code together for code understanding, implementation, debugging and testing
  • connects to Qwen and Qoder accounts, several model providers, and custom compatible endpoints
  • installs natively on macOS, Linux and Windows without a separate Node.js install
  • supports session resume and headless runs for scripts

Alibaba's announcement, covered by IT之家, frames it as a shared runtime other products can embed: an in process event bus today, with reasoning and tool execution kept separate so the design can later spread across several machines. In plain terms, Qwen Code is now the open source copy of the harness Qoder runs on. If that word is new to you, our guide to what an agent harness is explains why the harness matters as much as the model.

Two caveats from the 1.0 README: interfaces, settings and behavior may still change before it stabilizes, and the team is not accepting community pull requests until after December 2026. Issues are still welcome.

The release moves fast. The GitHub release is tagged 1.0.0-preview, but the official installer's channel manifest already pointed to 1.0.0-preview.1 when we checked on Nov 12. The preview changelog says that build fixes Qwen subscription accounts that could not use subscription models after sign in, and lets Qoder models make parallel tool calls.

Qwen Code 1.0 vs 0.27: what changed

Area0.27 stable (main branch)1.0 Preview
Installnpm (Node.js 22 or later), Homebrew, or standalone scriptNative installer or GitHub release archive, no Node.js
Updatesnpm or Homebrew, or qwen updateqwen update, plus qwen rollback to the previous version
Custom modelsmodelProviders grouped by protocol, plus security.auth.selectedTypeproviders entries with type, protocol, baseUrl and apiKey; models named provider/model
PermissionsApproval prompts, plan mode, sandbox commandFive modes: default, accept_edits, auto, dont_ask, bypass_permissions
Project .envSearched upward through parent foldersOnly <project>/.qwen/.env, else the user one
ExtrasWeb shell, daemon, chat channels, IDE plugins, desktop appTerminal CLI only, per the 1.x docs

Spoken plainly: 1.0 is a leaner terminal agent with a new install path, a new settings format and stricter permission controls. The 0.27 line still has more features around the terminal.

Install and updates

The 1.x installation guide uses a script hosted on Alibaba Cloud storage. You can also download the archive for your platform from the GitHub release. The Linux x64 archive we pulled was 68.2 MB and unpacked to a single 157.5 MB executable. qwen update --check looks for updates without installing them, and qwen rollback returns to the version that was active before the last update. Both only manage native installs. If you installed through npm or Homebrew you are on the 0.27 line, and the npm latest tag still pointed to 0.27.2 (Nov 10) when we checked.

Models and providers

The 1.0 authentication page lists six entries under /auth (also /connect or /login): Qoder, Qwen Account, QwenCloud, Alibaba Cloud Model Studio, other provider presets, and a custom provider. The Qwen Account option requires an active Plus or Pro plan; there is no free sign in tier in the current dialog. Custom endpoints go in a providers block in your settings file, documented on the models and providers page. Each provider has a type (openai-compatible or alibaba), a protocol (openai, openai-responses, anthropic or gemini), a base URL and an API key that can reference an environment variable. You then pick a model as gateway/your-model-id.

That is a different shape from the 0.27 modelProviders block in the stable authentication guide. Both versions read user settings from ~/.qwen/settings.json, so do not let one install rewrite the other's file. The 1.0 CLI has a --config-dir flag for exactly this: point the preview at its own folder while you test it.

Permissions and trust

The 1.x permissions guide adds five modes. accept_edits approves file edits but not arbitrary shell commands, auto uses a classifier where available, dont_ask denies anything that would need a new approval, and bypass_permissions (or --dangerously-skip-permissions) skips normal prompts. The docs say clearly that none of these is an operating system sandbox. If you know Claude Code, the flags will look familiar: --permission-mode, --allowed-tools, --disallowed-tools and --worktree all appeared in the help output we ran.

One detail matters for automation. Headless runs (-p, piped input, CI=true or GITHUB_ACTIONS=true) treat the working folder as trusted, so project settings, hooks and project MCP servers load without asking. When you point the agent at a repository you have not reviewed, add --setting-sources user. The headless guide also lists exit codes you can branch on in scripts: 0 success, 1 failed result, 41 authentication error, 42 input error, 52 configuration error and 130 cancelled.

What 1.0 does not have yet

The 1.x docs say they cover the terminal CLI. The 1.0 help output listed commands for MCP, plugins, skills, hooks, agents, update and rollback. The 0.27.2 help we ran also lists serve, daemon, channel, review, batch and sandbox commands, which back the web shell, the shared daemon, chat apps such as Telegram and WeChat, and scripted reviews. If your setup depends on any of those, or on the VS Code, JetBrains or Zed plugins, 1.0 is not a drop in replacement today.

What we saw when we ran Qwen Code on the box

We kept our hands on test small and free. On the stable 0.27.2 build, one headless prompt asking the agent to read a one line README produced three requests to our mock server, not two. The first carried a 42,417 character request body and only six tools (glob, grep search, read file, skill, tool call and tool search), so Qwen Code loads most tools on demand through search. The second sent the file contents back and got the answer. The third was a separate "managed memory extraction" subagent, a 26,448 character request with edit and write tools, whose instructions point it at ~/.qwen/memories and a per project memory folder.

With "memory": {"enableManagedAutoMemory": false} in ~/.qwen/settings.json, the same prompt made two requests. On a metered API key, that background call is worth knowing about; turn it off if you do not want it. The stable build also printed a warning that Node.js 24.22.1 can cause high temporary memory use during shell safety checks and suggested Node.js 26.12.1, which is one more reason the native 1.0 binary appeals.

On the 1.0 build we only ran --version and --help, so we cannot yet say whether the preview makes the same background call.

Qwen Code pricing and models

The CLI itself is free and open source. You pay for the model. With a Qwen Account you use your Plus or Pro subscription from chat.qwen.ai; QwenCloud and Model Studio offer pay as you go keys and plan based keys; other providers bill you directly. Qwen's docs do not publish one price for Qwen Code, so check the provider's own pricing page before you pick. Because both lines accept any OpenAI compatible endpoint, you can also point them at a local server such as Ollama's OpenAI compatible API and pay nothing per token, as long as the model handles tool calls well.

Should you switch to Qwen Code 1.0?

  • Try it now if you work only in the terminal, want a single binary without Node.js, or want fine grained permission modes for headless jobs. Use --config-dir and a test repository.
  • Stay on 0.27 if you use the IDE plugins, the desktop app, the web shell, chat channels or qwen review, or if a CI pipeline depends on current flags. The stable branch is still maintained.
  • Compare before you commit if you are choosing a terminal agent from scratch. OpenCode is the closest model agnostic rival, and our roundup of the best open source coding agents and our Claude Code vs Codex vs Gemini CLI comparison cover the main alternatives. Gemini CLI, the project Qwen Code was forked from, no longer has a free tier.

Whichever line you run, the same habits help: keep an AGENTS.md style instruction file in the repo, add skills for repeat tasks, and connect tools through MCP rather than broad shell access.

FAQ

Is Qwen Code 1.0 stable?

No. It is labeled Preview, and the README warns that interfaces, settings and behavior may change before it stabilizes. The stable 0.27 line is still maintained on the main branch, and the npm package still installs 0.27.

What is the difference between Qwen Code and Qoder?

Qoder is Alibaba's commercial agentic coding platform, with a desktop app, IDE, CLI and cloud agents, metered in credits. Qwen Code is the open source terminal agent. In 1.0, the harness built for Qoder moved into Qwen Code, and a Qoder account is one of the sign in options.

Is Qwen Code free?

The software is free under Apache 2.0. Model use is not automatically free: the Qwen Account route needs a Plus or Pro subscription, and API routes bill per use or by plan. A local model behind an OpenAI compatible server costs nothing per token.

Can I install Qwen Code 1.0 with npm?

Not at the moment. The npm and Homebrew packages install the 0.27 line. Version 1.0 Preview comes from the native installer script or the archives on the GitHub release page, and only native installs can use qwen update and qwen rollback.

Can I run 0.27 and 1.0 side by side?

Yes, with care. Both are called qwen and both use ~/.qwen by default, so check which executable your shell runs with qwen --version, and start the preview with --config-dir pointing at a separate folder so the two settings formats never share a file.