Skip to content
Technology

Alibaba's Qwen4-Max Beats Its Own Last Model

Alibaba's Qwen4-Max beat its last model on an independent score, but its own token-savings claim overstates what an outside lab found.

Edited by Axel4 min read · 1,030 words · 8 sources
A street-level view of Alibaba's Chaoyang Technology Park campus in Beijing
“Alibaba said Qwen4-Max needed about half as many tokens to beat its last model. An independent lab checked and found a smaller cut, 39 percent against the 50 Alibaba claimed. Does getting the score right excuse getting the savings wrong?”

Alibaba released Qwen4-Max on Nov. 16. The model beat its own last flagship by nine points on an independent test12. It still trails one Chinese rival. One of its biggest claims rests on Alibaba's own math alone4.

The Qwen team calls Qwen4-Max "your new AI teammate for work, creativity, and discovery"1. Artificial Analysis, an independent lab, tested the model the same day2. The two sides of this launch start right there: a company's own language, and an outside lab's number.

What Qwen4-Max Changes

Qwen4-Max runs on a design called mixture of experts. The design splits a model into smaller parts, each trained on a narrow slice of knowledge. Each question wakes only the parts it needs, leaving the rest of the network idle1.

The model holds 2.4 trillion parameters in total, and only 71 billion of them answer any single question6.

Qwen4-Max reads text and images in one window. The window holds close to a million tokens, the small pieces of text a model reads and writes12. Alibaba also cut the price. It charges $1.50 for a million input tokens and $4.50 for a million output tokens, down from $2.00 and $6.00 on the last model.

The catch sits in how a reader can use it. Qwen4-Max runs only through Alibaba's own cloud and apps today. The weights stayed closed, the files a company could otherwise download and run on its own servers6. A promise follows: smaller Qwen4 models will open later, on a date still unannounced1.

Qwen4-Max Clears One Rival, Trails the Other

Artificial Analysis grades frontier models the same way, so buyers can compare labs directly2. Qwen4-Max scored 43 on its Intelligence Index3.

That score sits nine points above Qwen3.8-Max's 34, and three points below Kimi K3.1 Pro's 46, the model Moonshot AI built2.

Qwen4-Max Jumps 9 Points to 43, Landing 3 Behind Kimi K3.1 Pro

Artificial Analysis Intelligence Index score, by model, as of Nov. 16, 2026. Higher is better.

Each column is one model's score on Artificial Analysis's Intelligence Index; a taller column scores higher, and the magenta column is Qwen4-Max.

Data through Nov. 16, 2026.

Sources: X (Artificial Analysis, @ArtificialAnlys) [2]; Artificial Analysis [3]. Chart by Kelly Dennis.

The numbers behind this chart
ItemValue
Kimi K3.1 Pro46
Qwen4-Max43
Tencent Hy435
Qwen3.8-Max34

Price tells a sharper story. Qwen4-Max costs $3.27 to run one Intelligence Index task. That is well under Qwen3.8-Max's $10.02, and 11 percent below what Kimi K3.1 Pro charges for the same task2.

A cheaper rival still beat it on price. Tencent's Hy4 model finishes the same task for $1.03, scoring 35 on the same index, eight points under Qwen4-Max2. Price and score pull in different directions once a buyer checks both.

Alibaba's Binjiang campus in Hangzhou, where its Qwen team builds its AI models
Alibaba's campus in Hangzhou's Binjiang district, photographed in 2021. The company's Qwen team, based in Hangzhou, released Qwen4-Max on Nov. 16, 2026.Charlie fong / Wikimedia Commons, CC BY-SA 4.0

One Gain Confirmed, Another Overstated

Alibaba's own launch post set the bar high. "Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation," the Qwen team wrote on X1.

Qwen@Alibaba_Qwen

📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!

Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.

Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over… pic.twitter.com/2Rlkoc6Dm3

Qwen's benchmark chart attached to its Qwen4-Max launch post
Open on X (opens a new tab)

Qwen's own launch post for Qwen4-Max on Nov. 16, 2026, the post the Qwen team used to make its efficiency and cost claims.Post by Qwen (@Alibaba_Qwen) on X, Nov. 15, 2026

Artificial Analysis ran the same comparison itself, outside Alibaba's walls. "Qwen4 Max uses 39% fewer output tokens per task than Qwen3.8 Max: ~105k output tokens per Intelligence Index task vs ~171k for Qwen3.8 Max," the lab wrote2.

Qwen4-Max's Measured Token Cut Trails Alibaba's Own Claim by 11 Points

Cut in output tokens per task vs. Qwen3.8-Max: Alibaba's claim against Artificial Analysis's own measurement

The blue dot is the token cut Alibaba advertised in its own launch post; the magenta dot is the cut Artificial Analysis measured running both models itself.

Alibaba's 50 percent figure comes from its own internal evaluation; Artificial Analysis measured the 39 percent figure running both models itself on its own Intelligence Index tasks.

Data through Nov. 16, 2026.

Sources: X (Qwen, @Alibaba_Qwen) [1]; X (Artificial Analysis, @ArtificialAnlys) [2]. Chart by Kelly Dennis.

The numbers behind this chart
ItemAlibaba's claimArtificial Analysis measured
Output tokens per task50%39%

A 39-point token cut is real, short of the 50 Alibaba advertised. Its cost claim holds up better. Artificial Analysis's own numbers show a 67 percent drop in price per task, past the 60 percent Alibaba promised2.

Jinhao Song, who covers model launches for OrcaRouter.ai, read Alibaba's own benchmark numbers the same week. "It is also a vendor number on a vendor benchmark, and it is labelled as such here because nothing independent reproduces it yet," Song wrote4.

The weights carry their own question mark. Mohammad Hassan Amer, writing for DEV Community, tested the access rules Alibaba set on its biggest model. "You cannot download it and run it on your own machine, the way you could with earlier Qwen models," Amer wrote5.

Nvidia Had Its Own Morning

Two more AI stories broke the same morning, both about one chipmaker apart from Alibaba. Both show how fast the chip race keeps moving.

Nvidia introduced the DGX Spark HPC Developer Kit at a supercomputing conference in Chicago on Nov. 167. Four of the units link together over Nvidia's own networking gear. Combined, they hold 512 gigabytes of memory. They can run models as large as 700 billion parameters, enough for most of what a research lab needs.

The Kraken supercomputer at Oak Ridge National Laboratory
The Kraken supercomputer at Oak Ridge National Laboratory in Tennessee. Nvidia's new DGX Spark kit packs similar reach into a four-unit desktop cluster unveiled Nov. 16, 2026.Daderot / Wikimedia Commons, Public domain

The US government moved the same day. It now requires a license first. A data center in Singapore, Thailand, Malaysia, Indonesia or Brazil needs that license before it hosts remote AI work. The rule targets computing tied to restricted buyers8.

Regulators built the rule to close a path Chinese firms used: renting chips abroad, skipping a direct buy at home. Operators of those overseas data centers now face a new compliance cost on business they already booked.

Read together, both stories point one way. Compute keeps landing on smaller machines, and the rules on who may rent it keep getting tighter.

What to do

Treat Qwen4-Max's independent score as real, and treat its internal numbers as a claim still in testing. Artificial Analysis measures a nine-point gain the same way across every lab, so that number travels23.

Ask any vendor pushing a token-savings figure who measured it. Qwen4-Max's own 50 percent figure came from Alibaba's internal evaluation alone12.

Price the task before you price the chat. Qwen4-Max beats Kimi K3.1 Pro on cost today, by roughly 11 percent per task2. Tencent's cheaper Hy4 model still wins on price alone.

By the numbers

  • 43: Qwen4-Max's score on Artificial Analysis's Intelligence Index, nine points above Qwen3.8-Max2.
  • 46: Kimi K3.1 Pro's score on the same index, three points above Qwen4-Max2.
  • $3.27: Qwen4-Max's cost per Intelligence Index task, against $10.02 for Qwen3.8-Max2.
  • 39 percent: the cut in output tokens per task Artificial Analysis measured, against Alibaba's own claim of up to 50 percent12.
  • 2.4 trillion: Qwen4-Max's total parameters, with 71 billion active on any one answer16.
  • $1.50 and $4.50: Qwen4-Max's price per million input and output tokens, down from $2.00 and $6.00 on Qwen3.8-Max1.

What to watch

Watch whether a second lab reruns these same numbers on its own hardware. Artificial Analysis already confirmed the score gain and the cost drop; the token-cut gap is the number still worth checking twice24.

Alibaba's promise on smaller, open Qwen4 models deserves the same scrutiny. It left the release date unannounced, and an unannounced date is easy to let slip1.

Sources

  1. Qwen, "Qwen4-Max is your new AI teammate for work, creativity, and discovery," post on X, Nov. 16, 2026, https://x.com/Alibaba_Qwen/status/2122148127520289151
  2. Artificial Analysis, "Alibaba has launched Qwen4 Max, gaining 9 points over Qwen3.8 Max on the Artificial Analysis Intelligence Index," post on X, Nov. 16, 2026, https://x.com/ArtificialAnlys/status/2122161629592953032
  3. "Qwen4 Max - Intelligence, Performance & Price Analysis," Artificial Analysis, Nov. 16, 2026, https://artificialanalysis.ai/models/qwen4-max
  4. jinhao.song, "Qwen4-Max Listed Elsewhere at Alibaba's Own Global Price," OrcaRouter.ai, Nov. 16, 2026, https://www.orcarouter.ai/blog/qwen4-max-day-one-third-party-api
  5. Mohammad Hassan Amer, "Qwen4-Max: Alibaba's Biggest Model Yet Is Cheap, and Closed," DEV Community, Nov. 16, 2026, https://dev.to/hassan_devs/qwen4-max-alibabas-biggest-model-yet-is-cheap-and-closed-59hd
  6. "Qwen4-Max: Features, Benchmarks, and Pricing," DataCamp, Nov. 16, 2026, https://www.datacamp.com/blog/qwen4-max
  7. "NVIDIA DGX Spark Developer Kit Turns a Desk Into a Supercomputing Lab," NVIDIA Blog, Nov. 16, 2026, https://blogs.nvidia.com/blog/dgx-spark-hpc-developer-kit-sc26/
  8. "US tightens AI chip export controls to deter Chinese companies from accessing compute remotely," TechCrunch, Nov. 16, 2026, https://techcrunch.com/2026/11/16/us-tightens-ai-chip-export-controls-to-deter-chinese-companies-from-accessing-compute-remotely/

Header photo: HoweyYuan / Wikimedia Commons, CC BY-SA 4.0

In the market

AlibabaBABA

Market value $225 billion

$90.62 −3.98% since the last close

NvidiaNVDA

Market value $3.77 trillion

$156.14 −4.12% since the last close

Prices from Yahoo Finance, read with yfinance, and delayed. This is a record of price, not investment advice.

Where this story happensChicago and Singapore
  • Chicago United States
  • Singapore Singapore

Map: Natural Earth land through world-atlas, drawn by Ai Lately.

Cite this piece

Kelly Dennis, "Alibaba's Qwen4-Max Beats Its Own Last Model," Ai Lately, Nov 17, 2026, https://ailately.com/articles/alibaba-qwen4-max-beats-last-model

Tags: Alibaba · Qwen4-Max · Artificial Analysis · Kimi K3.1 Pro · AI benchmarks

Related, lately

More

Keyboard

j k
Move through a list
Enter
Open the selected piece
/
Search, and browse the columns
g then r
Right Now
g then a
Articles
g then s
The Signal
g then o
Opinion
g then b
Analysis
g then p
People
?
This sheet
Esc
Close

AI Right Now is free with an account.The Citation Audit is free with an account.Skills are free with an account.Tutorials are free with an account.You have read your 3 free articles.

Every article and comic, AI Right Now, the Citation Audit, Skills and Tutorials.

Join Ai Lately for free

Join with Google or your email. Already a member? Sign in

Press anywhere else for the home page.