Back to Blog
    AI & AutomationNov 16, 2026

    Qwen4-Max Is Live: Alibaba's Flagship AI Out-Codes Claude — But You Can't Download It

    Amir

    Amir Farid

    7 min read

    Qwen4-Max Is Live: Alibaba's Flagship AI Out-Codes Claude — But You Can't Download It

    Qwen4-Max Is Live: Alibaba's Flagship AI Out-Codes Claude — But You Can't Download It

    Alibaba launched Qwen4-Max on 16 November 2026, the first release in the new Qwen4 family and the company's flagship AI model. It is a 2.4-trillion-parameter sparse mixture-of-experts model with 71 billion parameters active per token, a 1-million-token context window, and a headline price of $1.50 per million input tokens. The catch: unlike every Qwen model before it, you cannot download it.

    Alibaba says Qwen4-Max out-codes Claude on its chosen benchmarks — 75.7 on DeepSWE v1.1, ahead of Claude Fable 5.1 — and costs roughly a quarter less than the previous flagship, Qwen3.8-Max. Those are vendor claims, not independent results. Early independent rankings tell a cooler story.

    What launched on 16 November

    Alibaba announced the launch on Sunday 15 November, and the model went live the following day, after a quiet preview that had been running in Model Studio since 11 November. It arrives in the same week Tencent released Hunyuan 4 and Moonshot released Kimi K3.1 Pro, as The Straits Times reported.

    Its sparse mixture-of-experts design — 2.4 trillion parameters in total, only 71 billion active per token — keeps running costs down for a model this large. It takes multimodal input (text, image, video) and produces up to 393,216 tokens per run: 262,144 reasoning, 131,072 final answer.

    Where you can use it

    Available through QwenCloud, QwenWork, Qoder and QwenStudio. The Token Plan subscription that carried the preview runs until 25 November 2026. No download, no weights file, no self-hosting.

    The benchmark claims — and how to read them

    Developer testing Qwen4-Max on a laptop showing a glowing AI coding agent at work

    The headline number is a DeepSWE v1.1 score of 75.7, which Alibaba says puts Qwen4-Max ahead of Claude Fable 5.1 and Moonshot's Kimi K3.1 Pro. DeepSWE is built around real repository tasks, so a lead there matters — if the number holds up. TechThrum's launch report covers the full set of claims.

    On Work360, a benchmark for long-horizon office and agent tasks, Alibaba reports 52.69 against 47.74 for Qwen3.8-Max, with about 35% fewer tokens; on Terminal-Bench 4.0, roughly 48–49% fewer. Alibaba also reports a secure-code pass rate of 31.8% against 22.0% for Qwen3.8-Max.

    All of these figures are Alibaba's own numbers, published in its own launch material, and they have not been independently verified. Our GDPx agent benchmark coverage carries the same caution: no model cleared roughly 12% on long-horizon tasks.

    What the numbers do not prove

    A benchmark lead proves a model is good at that benchmark — not at your codebase or your team's way of working. Read Alibaba's claims as a reason to test, not a reason to switch. See our Claude Fable 5.2 coverage and Kimi K3.1 Pro analysis for the same lesson.

    What it costs

    Qwen4-Max is priced at $1.50 per million input tokens and $4.50 per million output tokens through the QwenCloud API. Alibaba says that is about 25% below Qwen3.8-Max list rates in Singapore. For heavy agent coding, where output tokens dominate the bill, that gap matters — and the claimed token efficiency compounds it.

    Price is only half the comparison. The honest question is not whether Qwen4-Max is cheaper per token but whether it completes your tasks at a similar success rate for a lower total cost. A model that costs half as much per token but needs twice as many attempts is not cheaper. Measure cost per completed task, not cost per token — our guide to AI agents for small businesses explains how to scope that pilot.

    The closed-weights shift

    Here is the part of the launch that matters most for the long term. Qwen is the most downloaded open-weight model family so far, according to Bloomberg. Qwen4-Max breaks that pattern: no weights, no download, no licence. Smaller Qwen4 models are promised soon, with no date, sizes or licence terms given.

    With Qwen4-Max, the flagship is API-only: your data goes to Alibaba's cloud and your pricing follows their list. It is the same trade-off as Claude and GPT — but it is new for Qwen, and deserves a deliberate decision.

    What to do if you planned to self-host

    Do not assume the smaller Qwen4 models will be open weight — Alibaba has promised them without committing to a licence. If self-hosting is a hard requirement, try genuinely open alternatives: our Mistral Large 4 open-weights coverage and our DeepSeek SuperAgent guide cover models you can actually run yourself.

    The honest counterpoint: independent rankings are cooler

    Agent Arena's live-session leaderboard, published 13 November and based on about 2.98 million sessions, puts Qwen4-Max (xHigh) in 14th place with a 7.46% task success rate. Claude Fable 5.2 sits first at 28.61%, and Claude Fable 5.1 is sixth at 14.10%.

    It predates the full launch and measures live sessions, not Alibaba's chosen benchmarks, so it is not a direct rebuttal — and every model on it scores far below what vendor charts imply. But the gap is large enough to matter. Our Cognition SWE-2 Flash coverage and our Ripple 65B preview show how fast this field is moving.

    Why the two pictures differ

    Vendor benchmarks use curated tasks; live leaderboards use whatever real users actually ask for. Trust neither chart on its own: run your own tasks, on your own code, and compare the results side by side.

    What your business should do this week

    Qwen4-Max is cheap enough and capable enough to earn a real evaluation, but not so proven that it earns an automatic switch.

    Step 1: pick five real tasks

    Choose five coding jobs your team actually did recently: a bug fix, a small feature, a refactor, a test-writing job and a code review. Use the original tickets, not toy examples — and include one task touching your automation workflow or RAG chatbot.

    Step 2: run them on your current model

    Run all five on the model you use today and record three numbers each: completed, tokens used, time taken. Our Gemini 4 Carbon coverage and Gemini 4 Hydrogen analysis are useful references if Gemini is your baseline.

    Step 3: run the same five on Qwen4-Max

    Repeat the five tasks on Qwen4-Max through QwenCloud, with the same prompts and success criteria. Record the same three numbers, plus the cost per completed task — judge it on that, not on price per token.

    Step 4: decide before 25 November

    Decide before the Token Plan window closes on 25 November. If the results are close, a hybrid setup — Qwen4-Max for routine coding, your current model for the hardest tasks — is a sensible middle ground. If you are planning a larger build, our guide to hiring an AI development company in London, our Dubai hiring guide and our UK custom software cost guide explain what to look for and how to budget it.

    Safety and the fine print

    Alibaba says security checks were built into training — code injection, memory-safety flaws, broken access controls — with safeguards against cybercrime and hazardous-materials misuse, and says internal testing found no significant high-risk capabilities. Those are Alibaba's own assurances, and they do not replace your own review. The Baidu Ernie 5.5 Flash preview and the Meituan LongCat-2.5 launch carry the same lesson: treat every new model as unproven until your own tests say otherwise.

    Frequently asked questions

    Is Qwen4-Max free to download?

    No. Qwen4-Max is not open weight: there is no download, no weights file and no self-hosting option. It is available only through Alibaba's cloud services — QwenCloud, QwenWork, Qoder and QwenStudio.

    How much does Qwen4-Max cost?

    Through the QwenCloud API it costs $1.50 per million input tokens and $4.50 per million output tokens — about 25% below Qwen3.8-Max list rates in Singapore, Alibaba says. The Token Plan subscription runs until 25 November 2026.

    Does Qwen4-Max really beat Claude at coding?

    On Alibaba's chosen benchmarks, yes: the company reports 75.7 on DeepSWE v1.1, ahead of Claude Fable 5.1. But those are Alibaba's own numbers, not independently verified. Early independent rankings are cooler — Agent Arena's live-session leaderboard puts Qwen4-Max 14th at 7.46% task success, behind Claude Fable 5.2 at 28.61%.

    Should my business switch to Qwen4-Max?

    Not on the launch announcement alone. Run five real coding tasks on Qwen4-Max and on your current model, compare cost per completed task, and decide before the Token Plan window closes on 25 November. If it matches your current model on quality and wins on cost per completed task, it is a genuine alternative.

    #Qwen4-Max#Alibaba#AI coding agents#coding benchmarks