Alibaba opened developer access to Qwen4-Max on November 11 — the first model in its Qwen4 generation. You have until November 25 to try it. After that, the preview model retires with no production path announced. That hard deadline, combined with zero published benchmark scores three days in, makes this one of the stranger flagship launches of 2026.
What Qwen4-Max Is
Qwen4-Max is Alibaba’s new multimodal reasoning flagship. It takes text, image, and video inputs and runs deep reasoning by default. Alibaba targets it at coding, frontend design, office productivity, and complex agentic workflows. The main new number: a 393,216-token output ceiling — three times what Qwen 3.8 Max could produce, and larger than most competitors. Context window is 1 million tokens, in line with Gemini 4 Hydrogen and GPT-6.1 Terra. The API is compatible with both OpenAI and Anthropic formats, so migrating existing code is straightforward.
Reasoning comes in three levels: low, medium, and xhigh, with xhigh on by default. One thing to budget for: the max_tokens limit covers both reasoning tokens and your final answer combined. Push xhigh on a long prompt and your effective output shrinks.
How to Access Qwen4-Max Preview
Access requires an Alibaba Cloud Token Plan subscription — from around $3 a month for new users. There is no pay-as-you-go option for the preview. Once subscribed, the API endpoint is token-plan.maas.qwencloudapi.com/v1 and the model ID is qwen4-max-preview. If you want OpenAI SDK compatibility without touching Alibaba’s endpoint directly, CometAPI offers an access path at their v1 endpoint.
from openai import OpenAI
client = OpenAI(
base_url="https://token-plan.maas.qwencloudapi.com/v1",
api_key="YOUR_TOKEN_PLAN_KEY"
)
response = client.chat.completions.create(
model="qwen4-max-preview",
messages=[{"role": "user", "content": "Your prompt"}],
extra_body={"enable_thinking": True, "thinking_budget": "xhigh"}
)The Catch: What Token Plan Bans
The Token Plan terms are restrictive in ways that matter for developers. Alibaba explicitly bans automation scripts, custom application backends, batch calls, and scheduled background tasks. The plan is for interactive use inside coding and agent tools only — Cursor, Claude Code, Codex, and similar. Sharing your key or running it from a CI pipeline risks getting the key suspended.
That means you cannot test Qwen4-Max the way you would test any other model. No automated evaluation harness. No side-by-side benchmark run. No integration test against your production pipeline. You get to poke it manually and that is it.
The Missing Benchmarks
Three days after launch, Alibaba has published no benchmark scores. No Artificial Analysis numbers. No HLE or SWE-Bench results. LMArena shows nothing. In 2026, every major model launch has had Artificial Analysis scores within 24 hours. Qwen4-Max is the exception, and the community has noticed. On Hacker News, one commenter noted: no benchmarks yet, and nobody they know actually runs Qwen Max. Without numbers, there is no basis for an adoption decision — just Alibaba’s self-reported claims about coding and reasoning improvements.
The November 25 Deadline
The preview ends November 25 at midnight UTC+8. The model retires. Alibaba has not announced a general availability date, a per-token price, or when the rest of the Qwen4 family — Qwen4-Plus, Qwen4-Flash, and Qwen4-27B open weights — will appear. Qwen4-27B weights were promised alongside the flagship cloud release. They are not here yet.
What that means practically: any code you write against qwen4-max-preview needs to change after November 25. Do not build anything on this endpoint. This is an evaluation window, not a launch.
What to Do Before November 25
If you are building Qwen-powered applications or evaluating models for long-context tasks, the 393K output tokens and 1M context are worth testing manually before the window closes. Video input is available if that is relevant to your use case. The OpenAI-compatible API makes it easy to swap in. Use it through Cursor or Claude Code — not through a script — to stay within Token Plan terms.
What you should not do: deploy it, build on it, or treat it as a representative benchmark for the full Qwen4 lineup. The preview model is frozen, the access rules prevent real evaluation, and the missing benchmarks make honest comparison with Gemini 4 Hydrogen or GPT-6.1 Terra impossible for now. Wait for GA, independent scores, and an actual per-token price before making any decisions.
