A generated title card headed 'Qwen4-Max Is Free for Five Days' showing a window running 18 to 23 November 2026, three labelled parties — Alibaba Qwen, Nous Research and Inferact — and a footer noting that the model's metered rate card is unchanged.
Guides & Insights

Qwen4-Max Is Free for Five Days on Nous Portal — and the Serving Partner Is the Story

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

At 19:25 UTC on 18 November 2026, Nous Research posted that Qwen4-Max, Alibaba's new flagship, was free in Nous Portal for five days in partnership with Inferact and the Qwen team. Seventy-eight minutes later, at 20:43 UTC, Inferact posted its own version of the same announcement: five days of free tokens on Qwen4-Max, served on Inferact Platform and on vLLM. Two posts, no blog entry, no technical write-up, no price table. Qwen4-Max itself did not change in any way that afternoon — it went generally available at its metered rate on 16 November, and that rate is still the rate. What changed is that the model now has a publicly named serving stack, and for the next few days anyone can go and push work through it for nothing.

That distinction is the whole article. A free window is not a price cut, and a partnership announcement is not a benchmark. But the second sentence of Inferact's post is the part worth reading twice: this is a model Alibaba has never published a checkpoint for, being served on an open-source inference engine by a company that did not build it. Whatever else the promotion is, it is the first public evidence of how Qwen4-Max behaves outside Alibaba's own deployment.

What the two posts actually say — and what they leave out

Nous Research's post is the substantive one. It calls Qwen4-Max "the new flagship model from Alibaba," says it is "built for long-horizon work, with native vision and 1M context," and points readers at the Portal. Inferact's post adds the operational detail: the free tokens are served on Inferact Platform and on vLLM, and the word it uses for what it is giving away is "tokens," not "requests."

Neither post states a token cap, a per-account limit, a rate limit, or an end timestamp. The Portal's model list currently shows Qwen4-Max under Free Models as qwen/qwen4-max:free at a price of FREE, alongside a note that prices are updated every minute from the live catalogue — which tells you the listing is real and live, and tells you nothing about how much of it you can consume. If the clock started with the posts on the evening of 18 November, the window runs to roughly 23 November. Treat that as an inference, not a published deadline.

So the honest summary of the offer is: five days of metered access to a model that otherwise costs real money, with the meter switched off and the ceiling undisclosed. If you have been waiting for a reason to evaluate Qwen4-Max before committing a production path to it, this is the cheapest reason you will get.

Why "served on Inferact and vLLM" is the interesting half

Inferact is the company founded by the creators and maintainers of vLLM, and it sells a production inference platform built on top of that engine. Its pitch is day-zero model support, custom kernels, autoscaling and metering, either on your own GPUs or on theirs. Naming it as the serving partner is therefore not a hosting footnote; it is a statement that Qwen4-Max can be run on a serving stack that is not Alibaba's, by engineers who tune kernels for a living.

That raises an obvious question, because Qwen4-Max has no public weights. Alibaba publishes no checkpoint for it — the Model Studio documentation lists fine-tuning, model deployment and batch inference as unsupported, and there is no Hugging Face repository to download. A model you cannot download cannot, in the ordinary sense, be served on vLLM by a third party. The only way both statements are true at once is that Alibaba gave Inferact the weights under a commercial arrangement and Inferact tuned and packaged them. Which means this partnership is the closest thing to a public look at the model's serving behaviour that exists — a look at the engine, not at the file.

For a team deciding whether to build on Qwen4-Max, that is more useful than it sounds. The questions that decide whether a long-horizon agent is viable are rarely about the benchmark row. They are about time to first token, throughput under concurrency, how the thinking budget is actually enforced, and whether the engine falls over when a 900,000-token request arrives. Those are exactly the questions a second serving stack lets you ask, and exactly the questions Alibaba's launch material does not answer.

Screenshot of the Nous Research Portal model list captured 19 November 2026, showing 'Qwen: Qwen4 Max' listed under Free Models with the model ID qwen/qwen4-max:free and a Portal price of FREE, above a note that prices are updated every minute from the live catalogue.

The meter is off for five days; the rate card is not going anywhere

It is worth restating what the model costs once the window shuts, because the free period is designed to make you forget it. In Singapore, Alibaba lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output, with cached input at $0.075 and cache writes at $1.875 for the five-minute window or $3.00 for the hour. The Global deployment scopes — Beijing, Hong Kong, Frankfurt, Japan and US Virginia — run slightly lower at $1.413 in, $4.24 out and $0.071 for a cache read. There is one price tier all the way to the context limit: no long-context surcharge, no batch discount to hunt for, because batch inference is not offered on this model at all.

Then there is the part that decides your bill more than the per-token rate does. Thinking is on by default. The default reasoning effort is high, the default thinking budget is 131,072 tokens, and the combined ceiling for reasoning plus visible answer is 262,144 tokens. Non-thinking output is capped at 131,072. A long-horizon agent run that thinks hard and answers long is therefore spending output tokens for minutes at a time, and output is the expensive side of the meter by a factor of three.

Worked through on a real shape of job: 40,000 input tokens and 120,000 combined thinking-plus-answer tokens costs $0.06 of input and $0.54 of output, so about 60 cents a run at Singapore rates. Five hundred such runs is $300. That is the bill the free window is holding off — and it is also the number to keep in front of you when the window closes and a prototype quietly becomes a nightly job.

A generated price card headed 'Qwen4-Max: what five days is worth', listing Singapore rates of $1.50 per million input tokens and $4.50 per million output tokens, cache reads at $0.075, a worked example of 40,000 input plus 120,000 thinking-and-answer tokens costing about $0.60 per run, and a five-day window with no published token cap.

What you can actually finish in five days

Here the independent numbers get uncomfortable in a useful way. Artificial Analysis, running its Intelligence Index v5.0.1, measures Qwen4-Max at 43 and records the model at 31.3 minutes per index task against a median of 12.4 minutes for comparable models. Output speed is 56 tokens per second, and time to first answer token is 146 seconds. It generated 127 million output tokens across the index run, against a median of 65.1 million — the model is verbose, and the verbosity is priced at the output rate.

Those two measurements combine into a planning constraint nobody has written down. Five days is 7,200 minutes. At 31.3 minutes per task, a single serial lane of work completes roughly 230 tasks. Four parallel lanes get you to about 920. If your evaluation plan is "we'll try a few prompts," the window is enormous. If your plan is a 2,000-case sweep of a long-horizon agent, five days across one account will not finish it even with the meter off — and nothing in either announcement says how much concurrency the free tier will tolerate.

The same arithmetic explains why the free window is genuinely valuable rather than merely generous. What you cannot buy cheaply is calendar time with a slow model. A week of unmetered access converts directly into a week of latency data, token-usage data and failure-mode data at zero marginal cost, on a model whose per-task compute cost is $3.27 on the independent board. That is the experiment worth running, and it has to be designed before the 23rd, not during it.

Screenshot of the Artificial Analysis page for Qwen4 Max captured 19 November 2026, showing an Intelligence Index v5.0.1 score of 43 at rank 51 of 96, time to first answer token of 146.14 seconds, output speed of 56 tokens per second, 31.3 minutes per index task against a median of 12.4, $3.27 cost per task, and $1.50/$4.50 per million token pricing.

The one number to measure before the window closes

If you take one measurement out of the five days, make it output tokens per finished task on your own workload, at your own reasoning setting, rather than on a leaderboard. Everything about this model's economics hangs on it. The vendor's own launch framing — unreproduced by anyone outside Alibaba — is that Qwen4-Max outperforms Qwen3.8-Max while using up to about 50% fewer tokens, with average cost per task falling from $4.43 to $1.67. The independent board's reading is different in emphasis: 43 on the index, and 127 million output tokens over the run. Half the tokens of the previous generation can still be a lot of tokens.

The reason to measure it yourself, this week, is that the free window is the only period in which you can run the token-hungry version of the experiment without a budget conversation. Turn thinking to low and to xhigh on the same task set; log the combined thinking-plus-answer total against the 262,144-token ceiling; watch what happens when a request approaches the 983,616-token thinking-mode input limit. Those are the numbers that decide whether Qwen4-Max is a nightly batch job or a per-request product feature, and they cost nothing to collect until the 23rd.

One practical warning about the reasoning setting while you are at it. Qwen4-Max exposes three effort levels — low, high and xhigh — and an OpenAI-shaped client that asks for medium or max gets folded into high or xhigh rather than rejected. A harness tuned against a five-level model will silently spend more than you planned, and the free window is precisely where you would fail to notice, because the invoice reads zero either way.

Where this leaves you on the 24th

When the tokens stop being free, nothing about the model will have changed: $1.50 and $4.50 in Singapore, 1M-token context, native image and video input, three reasoning levels with quiet remapping, and a per-task compute cost on the independent board that is high because the model thinks for half an hour. What will have changed is what you know about it. Teams that spend the window measuring token burn and tail latency will come out with a decision. Teams that spend it chatting will come out with an impression, and impressions of a 31-minute-per-task reasoning model are the least reliable kind.

To be straight about where we sit: OrcaRouter does not route Qwen4-Max, so nothing here is a pitch for our endpoint, and the free tokens are on Nous Portal and Inferact's platform, not ours. If the evaluation leads you to a paid deployment, our catalogue is one key across 200+ models with provider list pricing passed through at 0% markup, which means a vendor price change on a model we do carry is live the same day rather than at the next contract renewal. That matters more than usual for a model this expensive to run: when the rate card is the product, the difference between list price and marked-up price is the whole margin.