- Published on
Qwen4-Max is out, and it is not open weight. That is the story.
- Authors

- Name
- AbnAsia.org
- @steven_n_t
Listen to this article · Qwen4-Max is out, and it is not open weight. That is the story.
0:00 / 0:00
Watch this article · Qwen4-Max is out, and it is not open weight. That is the story.
Alibaba released Qwen4-Max at 02:01 UTC on 16 November, five days after a preview. It is a 2.4-trillion-parameter mixture-of-experts model with native vision, a one-million-token context window, and a public price: 4.50 per million output, $0.075 for implicit caching.
It is also not open weight. Alibaba's own post says what comes next: "the open weights of smaller Qwen4 models will be released. Stay tuned." No date, no sizes, no licence. Bloomberg reports that the Max model itself is closed.
That single line is the whole story, and it is not the story most of the timeline told.
What Alibaba claims
The company says Qwen4-Max beats its predecessor, Qwen3.8-Max, "with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation". Bloomberg's report adds the sharper claim: that it beats Fable 5.1 and Kimi K3.1 Pro on coding benchmarks.
Two things are true about that sentence at once. It is a big claim — Fable 5.1 is an Anthropic frontier model — and it is a claim made by the vendor, on the vendor's benchmark table, with no independent test behind it yet.
The table Alibaba published is worth reading rather than skimming, because it does not say what the headlines said. Across the twelve benchmarks, Qwen4-Max leads most and loses several: on Chartography, a visual-understanding test, GPT-6 Astra scores 71.1 and Qwen4-Max 54.2. On AgentArena, Qwen4-Max's 7.5 sits below its own predecessor's 18.2.
The price is the real product
For anyone building, the number that matters is not the benchmark row. It is 4.50 out, against 40 for Claude Fable 5.2, which shipped on 9 November and currently tops the Vals Index at 71.31%.
That is roughly a fifth of the price for a model that is competitive on the tasks this desk cares about — code, documents, charts. For an agent that loops all day, the cost per finished task decides what you can afford to attempt, and Qwen's own framing is exactly that: cost per task, not price per token.
So is the gatekeeping gone?
No — and this is where the excitement of the last 24 hours outran the facts. An open-weight model you can download is a different object from a hosted model you can call. Qwen4-Max at 2.4 trillion parameters was never going to run on a desk, open or not; the weights that would matter for that are the small ones, and those are the ones Alibaba has only promised.
What actually shipped on 16 November is an API, a price, and a licence decision. The capability arrived; the distribution did not.
11 November
Qwen4-Max preview opens on Alibaba's Token Plan, 1M context, no benchmarks, no price.
16 November
Release: API live on QwenCloud, 2.4T MoE, $1.50/$4.50, native vision.
Not yet
Open weights for smaller Qwen4 models, promised with no date, no sizes, no licence terms.
When they land
That is the day a person with a workstation can run the family locally. It has not happened.
What this means if you are building
Use it. At a fifth of the frontier price with a million-token context, Qwen4-Max is worth wiring into the boring, high-volume work today: document processing, code review, chart reading, long-running agents.
But price your architecture against the licence, not the launch post. If your product assumes you can download the weights, you cannot — not yet, and not for this model. Keep a second provider in the config, keep the model ID in one place, and treat the promised open release as a date to watch rather than a fact to build on.
The advice that survives the hype is the unglamorous version: learn the tools that exist this week, keep your costs measured per finished task, and do not confuse a company's frontier with your own capability. The frontier is rented. What you can run is yours.
Source: Qwen's release post on X, 16 November 2026, 02:01 UTC, and the benchmark chart published with it; Bloomberg's report on the licence as summarised by Techmeme; the OpenRouter listing for qwen/qwen4-max read on 16 November; Claude Fable 5.2 figures from Anthropic's launch material and the Vals Index entry dated 9 November 2026. No independent benchmark run of Qwen4-Max was available at publication, and no Qwen4 weights were listed in Qwen's Hugging Face organisation. The analysis is ours.
Contact: [email protected]