Alibaba’s Qwen team has published Qwen Code 1.0 Preview, an open-source terminal coding agent that now claims feature parity with Anthropic’s Claude Code on nearly every row of a comparison table, and beats it on four: open-source models and framework, multi-protocol support, an “Agent Arena” for multi-model head-to-head runs, and daemon mode. The repository’s own capability matrix puts a checkmark in both columns for SubAgents, Agent Teams, dynamic workflows, Auto-Memory, Auto-Skills, hooks, MCP, Plan Mode, LSP integration, sandboxing, Git worktrees, computer use, IDE plugins for VS Code, JetBrains and Zed, SDKs, headless mode and session management. That is a lot of checkmarks to concede.
The frankness is the first surprise. Most competing agent vendors write comparison pages that flatter themselves. Qwen’s page states plainly: “If you know Claude Code, you already know Qwen Code.” Then it lists the rows where Claude Code has no equivalent, and the gaps are structural rather than cosmetic.
What the parity table actually concedes
Anthropic’s Claude Code is closed. Its model is closed. Its provider list is Anthropic. Qwen Code ships the framework and the Qwen models as open source, and the README says they “evolve together.” It accepts OpenAI, Anthropic, Gemini and Qwen APIs, plus any third-party provider or local model via Ollama or vLLM, switchable at runtime. That last clause matters more than the rest of the table combined. A team that standardizes on Qwen Code is not locked to Alibaba’s inference bill; it can point the same harness at a self-hosted Qwen checkpoint, a Gemini endpoint, or an Anthropic key on a per-task basis.
The Agent Arena is the second interesting row. Running multiple models head-to-head on the same task inside one harness is the kind of feature that only makes sense for a vendor that does not depend on a single model winning. Anthropic has no incentive to build it. Alibaba does, because Qwen’s models need to be seen beating the alternatives on real work, not benchmark tables.
The SWE-bench number needs a caveat
The repo reports that seven Qwen Code versions scored 76.40% to 77.80% on average on SWE-bench Verified, “using the same Qwen 3.7 Max model (500 tasks, three trials per version).” Read that sentence twice. The spread across seven versions of the harness is 1.4 points, which is a tight band and suggests the harness is not the variable doing the work. The model is. Qwen 3.7 Max is doing the heavy lifting; the agent scaffolding contributes a consistent, modest amount.
{/* TODO: verify the Qwen 3.7 Max model name and release date against Alibaba’s official model documentation — the README references it but Tessera has not confirmed a public model card */}
That is not a knock on Qwen Code. It is the honest reading of any SWE-bench Verified score above 75% in late 2026: the benchmark is saturating and the differences between top harnesses are shrinking. The interesting question is no longer which agent scores highest. It is which agent runs cheapest, on whose hardware, under whose terms.
Daemon mode is the real differentiator
Buried under the feature table is qwen serve, a long-lived service that exposes an HTTP plus SSE API on http://127.0.0.1:4170 by default, supports multiple workspaces, and lets browser, editor, SDK and automation clients share agent sessions instead of each spawning a process. There is a Web Shell, a background daemon with qwen daemon start, status, logs -f and stop, and a documented warning that the daemon “listens on loopback without bearer authentication; any local process that can reach it can use the full API, including code execution.” Set QWEN_SERVER_TOKEN and --require-auth on shared hosts, the README says.
That warning is unusually direct for a vendor README, and it points at the architecture Qwen is betting on. A coding agent that is a process you invoke is a toy. A coding agent that is a service other software calls is infrastructure. The daemon, the TypeScript, Python and Java SDKs, the ACP-based qwen-code-claw skill that lets Claude Code or Codex delegate coding tasks to Qwen Code, and the headless qwen -p mode all push in the same direction: Qwen Code as a callable backend, not a chat window.
The daemon turns Qwen Code from a tool you run into a service other agents call. That is the row Anthropic has not shipped.
The self-hosting claim deserves scrutiny
The README’s tip says Qwen Code “is actively iterating on itself, using its own agent and models to file issues, submit PRs, review code, and run tests.” That is a strong claim. It is also the kind of claim that is easy to make and hard to verify from a README. The repository does have a visible skill directory (.qwen/skills/qwen-code-claw/SKILL.md) and a contributing guide, but the proportion of merged PRs authored by the agent is not stated.
{/* TODO: verify the self-iteration claim — check the repo’s commit history and PR authors for agent-generated contributions; the README asserts it without a figure */}
If true, it is the most consequential thing on the page. An agent harness that improves itself on real issues is a different category of product from one that is merely good at SWE-bench. If it is aspirational, the README is doing marketing.
What this means for builders
Three things are now table stakes for a coding agent in late 2026: multi-provider runtime, an SDK, and a daemon or server mode. Qwen Code ships all three in an open-source package that installs via npm install -g @qwen-code/qwen-code@latest on Node 22+, Homebrew, or a curl pipe. The install surface alone is broader than most competitors.
The harder question is trust. A daemon that can execute code, listening on loopback without auth by default, installed via a script hosted on an Alibaba OSS bucket in Hangzhou, is a supply-chain and security decision, not just a developer-experience one. The README’s own warning about QWEN_SERVER_TOKEN is a signal that the team knows this. Enterprises will want to see the threat model before they let qwen serve run on a shared build host.
The broader shift is that the coding-agent market is converging on a common feature set faster than anyone expected. Claude Code defined the shape. Qwen Code, Gemini CLI (which Qwen Code was originally based on, at v0.8.2, before the team stopped syncing upstream at v0.1), Codex and the rest are filling it in. The differentiators left are price, model quality, and openness. Alibaba is competing on the third and betting the first two follow.
Whether that bet pays depends on whether Qwen 3.7 Max stays competitive with the frontier. The 76.40% to 77.80% SWE-bench range says it is close. Close is not the same as ahead.