Confirmed 16 Nov 2026 · Labs A one-minute read

Alibaba's new flagship went on sale, and it says the same work costs 58% less.

Qwen4-Max went live on Alibaba's paid API on November 16. The caveat: Alibaba says the 58% comes from Work360, a benchmark it built and scores itself.

Until now

Frontier gains come from scaling up: more parameters active per token.

Now

Alibaba cut active parameters and says capability rose. Architecture is doing the work.

Why it matters

Cheaper tokens are a price war. Fewer active parameters at higher capability is an engineering claim, and if it holds, the cost of running a frontier model falls for everyone who copies the architecture. Alibaba will not release the weights, so the only way to test it is to pay.

Established

  • qwen4-max appears in Alibaba's own API documentation and requires its multimodal interface.
  • Alibaba published per-task cost figures against its own Work360 benchmark.
  • The weights are not released. Smaller Qwen4 models are promised later.

Not established

  • The parameter count. Alibaba has not published it; outlets report 2.4 trillion.
  • The 58% cut. It rests on a benchmark Alibaba built and scores.
  • That it leads. Artificial Analysis and LMArena have not ranked it.
The five-minute readThe record, the timeline, the weight and 5 sources

The record

Alibaba's documentation at help.aliyun.com, read directly, names qwen4-max in its official code examples for the OpenAI-compatible Chat Completions and Responses endpoints and the DashScope interface, and states that qwen4-max must use the multimodal interface, so text-interface code raises an error. The same site's model catalogue still lists only qwen4-max-preview, so the catalogue and the API documentation disagree. Everything that follows is reported. Reuters, datelined Beijing November 15 and filed 04:29 Universal Time on November 16, says Alibaba unveiled Qwen4-Max as its new flagship, and says training added security checks alongside coding tests that reduce code injection, memory-safety flaws and broken access controls, with strengthened safeguards against cybercrime and hazardous materials and no significant high-risk capabilities found in internal testing. IT Home, carried by Sina Finance at 10:07 Beijing time on November 16, reports a sparse mixture-of-experts backbone of about 2.4 trillion parameters, 71 billion active per token against 95 billion for Qwen3.8-Max, a 328 billion parameter n-gram embedding at the second layer, native visual understanding and a 1 million token context. Three of every four attention layers use Gated DeltaNet; the fourth uses Qwen Sparse Attention. On Work360, Alibaba's own framework covering 26 primary domains, 182 secondary subdomains and 721 tertiary specialties, it reports 52.69 against 47.74, token use down about 24% and 35% at two reasoning settings and about 48% and 49% on Terminal-Bench 4.0, and average cost per task down from 26.58 yuan to 11.16 yuan, a 58% fall against a headline claiming over 60%. Last Secure Code Bench is reported at 31.8% against 22.0%. BigGo Finance, filed 03:35 Universal Time, gives international prices of $1.50 per million input tokens and $4.50 output, implicit cache input at $0.075, 70% below the prior generation, and sixth place in CodeArena. A limited preview opened November 11; the Qwen blog post is dated November 15; paid availability is reported from November 16. Alibaba says smaller and mid-sized Qwen4 models will be open-sourced in stages with weights published, naming no dates. Not established: the weights, the licence, the full benchmark table, and any independent ranking.

How it got here

  1. Aug 17

    Alibaba releases Qwen3.8-Max, the predecessor, with 95 billion active parameters per token, and publishes weights for the Qwen3.8 flagship on its international blog.

  2. Sep 20

    Alibaba's Qwen team releases an image model and removes the right to use it commercially, carded in Edition 028. Two days later, in Edition 030, Alibaba says its model has been improving itself for a month and publishes the score.

  3. Oct 31

    Apple confirms its China AI features will run on Alibaba's Qwen model from November 4, carded in Edition 068. Qwen becomes the assistant layer for iPhones in the largest smartphone market.

  4. Nov 11

    A limited preview of qwen4-max opens. Alibaba's own model catalogue still lists only qwen4-max-preview today, 5 days later.

  5. Nov 15

    Huawei Cloud lists openPangu-2.1-Pro, calls it open source, and publishes no weights, no parameter count and no technical report, while retiring the genuinely open 2.0 models. Edition 095. The Qwen blog post carries the same date.

Who has to react

DeepSeek, Moonshot and Zhipu, whose open weights are the alternative; Huawei Cloud, one day into the same no-download position; Apple, whose China features run on Qwen; the Chinese chip makers serving this much inference; Artificial Analysis and LMArena, which have not ranked it; and every developer now choosing between a cheaper closed Chinese model and a slower open one.

What to watch

  • Would strengthen itArtificial Analysis or LMArena ranks Qwen4-Max at or above Qwen3.8-Max while Alibaba's published international price stays at or below $1.50 per million input tokens, by March 31, 2027.
  • Would weaken itAlibaba publishes downloadable weights for any Qwen4 model under a licence permitting commercial use, by March 31, 2027.
  • Would weaken itAn independent evaluation places Qwen4-Max below Qwen3.8-Max on a benchmark Alibaba did not build, or Alibaba withdraws the per-task cost claim, by June 30, 2027.

Why it weighs this much

Signal strength
CostMedium
CommitmentHigh
ReversibilityLow
PersistenceYear

Cost is medium. Training a flagship at this scale is expensive, but Alibaba has done it repeatedly and this is scheduled, not a bet. Commitment is high: prices are published, the model sits in Alibaba's own API documentation, and a flagship cannot be quietly withdrawn once enterprises build on it. Reversibility is low for the same reason, and because the architecture, linear attention on 3 of every 4 layers, is baked into a trained model, not a switch. Persistence is a year: the next flagship resets the numbers, and the promised open small models could change what open means here again.

The words

Mixture of experts
A model split into many sub-networks where only a few run per token. Total size can be huge while the cost of each answer stays closer to the active part.
Active parameters
The share of a model actually used to produce one token. It sets the compute bill, which is why cutting it while raising scores is the claim that matters.
Gated DeltaNet
An attention alternative that compresses past text into a fixed-size state instead of comparing every token with every other. Long inputs get cheaper, at some cost in recall.
Open weights
Publishing the trained numbers so anyone can run or change the model offline. Without them, a model called open source can only be rented.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?