Qwen4-Max Preview packs a 1M-token context and agent-first coding claims, but Alibaba shipped it with no benchmarks and a hard November 25 shutoff. Here's what's real.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Alibaba put Qwen4-Max into preview on November 11, 2026, and it disappears on November 25. That's not a typo or a soft "early access" window. The Token Plan subscription that gates access to the model gets retired at 24:00 China Standard Time on the 25th, and Alibaba has said the model goes with it. If you want to see what Qwen4 can actually do before it's gone, you've got about two weeks from the time most people hear about this.
Here's the weird part: Alibaba shipped this without a benchmark table, without a model card, and without open weights. For a lab that's been publishing exhaustive spec sheets for every Qwen release since 2.5, that silence is the real story.
Qwen4-Max Preview is the first public model in Alibaba's Qwen4 line, built on what the company describes as a new MoE architecture with native vision support. It handles text, images, and video, and ships with a 1M-token context window, big enough to drop an entire mid-sized codebase into one prompt and still have room for the conversation.
Max output sits at 393,216 tokens, split between the model's reasoning and its final answer. Thinking is on by default, with low, medium, and xhigh effort settings you can dial depending on whether you want a fast answer or a slow, careful one.
Alibaba's own numbers (not independently verified) claim Qwen4-Max beats its predecessor, Qwen3.8-Max, using roughly 50% fewer tokens per task, which the company says cuts average task cost by more than 60%. Take that with a grain of salt: it's Alibaba grading its own homework, and there's no public methodology behind it.
Alibaba is pitching this release at coding, office work, front-end builds, and what it calls "long-horizon agent tasks" (the multi-step, tool-calling workflows that drive things like autonomous PR reviews or multi-file refactors). The model is OpenAI- and Anthropic-API compatible, so it plugs into Claude Code, Cursor, and Cline without custom glue code. Alibaba is also pushing it through its own tools: QwenWork, Qoder, the Qwen app, and QwenStudio all got early access the same day.
Function calling, web search, structured JSON output, and built-in tool support are all there. In practice that means an agent running Qwen4-Max can search the web, call a tool, read the result, and keep going without you babysitting every step: the stuff that makes "agentic" more than a marketing word, assuming it holds up outside Alibaba's demos.
You can't just pay per token. There's no pay-as-you-go pricing for Qwen4-Max. Access runs exclusively through Alibaba's Token Plan subscription, individual or team. You pick a tier, get a dedicated API key, and point your client at https://token-plan.maas.qwencloudapi.com with the model ID qwen4-max-preview.
The tiers break down like this:
| Plan | Price | Credits | Intro rate |
|---|---|---|---|
| Lite | $6/month | 20,000 | $3 |
| Essential | $10/month | 25,000 | $5 |
| Standard | $18/month | 100,000 | — |
| Pro | $43/month | 200,000 | — |
The catch: the plan's terms explicitly ban bulk processing by automation scripts, using it as a permanent backend for an app, or embedding it in non-interactive batch pipelines. This is built for a human sitting in an IDE or chat window clicking "run," not for a production service. So if you were hoping to swap Qwen4-Max into your existing agent stack and let it grind overnight, read the fine print first.
This is the part that should make you pause before getting excited. Alibaba shipped Qwen4-Max with no benchmark table, no parameter count, and no model card (a sharp break from how it's launched every recent Qwen model). The only numbers floating around (Terminal-Bench 2.1 at 86.6, SWE-bench Pro at 67.7) actually belong to the previous model, Qwen3.8-Max, not this one. People are citing them as if they're Qwen4-Max's scores. They're not.
Independent testing is filling the gap, and it's mixed. Artificial Analysis put Qwen4-Max's Intelligence Index at 43: behind GPT-6.1 Sol's 46 on medium effort (55 at max), and roughly tied with Kimi 3.1. On its own, "roughly tied with Kimi 3.1" isn't a disaster. It's just not what "new architecture, 1M context, agent-first" usually implies on launch day.
One independent coding benchmark (GameFeel RE) actually scored Qwen4-Max lower than Qwen3.8-Max: 10.3 versus 12.1. That's a regression on at least one third-party test, which cuts against Alibaba's "beats the previous model" framing.
There's a theory going around Hacker News worth mentioning, even if you can't verify it: some commenters think Qwen4-Max underperforms because Anthropic has gotten better at blocking distillation (the practice of training a new model partly on a frontier model's outputs). If that's true, it would explain why a brand-new architecture landed in the middle of the pack instead of at the top. I can't confirm that from here, but it tracks with the pattern of Chinese labs posting smaller year-over-year jumps as the big three keep hardening their APIs against scraping.
Honestly, no: not for anything that needs to still work after November 25. This is a two-week test drive, not infrastructure. The usage restrictions in the Token Plan terms already rule out production use, and the hard retirement date means anything you wire up today needs a fallback plan by the 25th regardless.
What it's actually good for: kicking the tires. If you're curious whether a 1M-token context changes how you work through a large repo, or whether the agentic tool-calling loop is smoother than what you're using now, $3-6 for two weeks is a cheap way to find out. Just don't point anything that matters at it.
I'd also wait for the open-weight variants. Most of the excitement in the Qwen community right now isn't actually about Qwen4-Max: it's about the smaller 27B and Flash versions Alibaba is expected to release with open weights, the same pattern it followed with Qwen3.8-27B. Those are the ones people will actually self-host and compare honestly, without a subscription meter running or a countdown clock attached.
Qwen3.8-Max, covered here when it launched, set the baseline numbers everyone's now misattributing to Qwen4-Max. On paper, the new model is supposed to beat it using fewer tokens per task. In practice, the one independent benchmark that directly compared the two (GameFeel RE) had Qwen3.8-Max winning.
Against the rest of the field, Qwen4-Max lands below GPT-6.1 Terra and GPT-6.1 Sol on the Artificial Analysis Intelligence Index, and in the same range as Kimi 3.1 (a model we've already compared against Claude for frontend work). If you're choosing a daily driver for agentic coding today, none of this moves Qwen4-Max to the top of that list. It moves it into "worth a two-week look," which, to be fair to Alibaba, is exactly what a preview is supposed to be.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Google Antigravity's /teamwork needs a paid plan and can flood your folder with files. Claude Code subagents are cheaper but need manual orchestration. Here's which wins for long projects.
Pinky is a free open-source MCP gateway from DIYgod that lets Claude Code or Codex control your real iPhone or Android phone with a human approval gate for risky actions.
Expo's new Exact framework runs one Contract app natively on web, macOS, iOS and Linux in a ~20KB runtime, built for AI coding agents to write, run and test it.
Comments