metir
metir
Docs
Download on App StoreGet it on Google PlayLog inSign up
Back to Blog
Alibaba
Qwen
Open-Weight Models
AI Pricing
LLM Benchmarks

Alibaba Qwen4-Max: Benchmarks, Price and Closed Weights

Alibaba opened Qwen4-Max in preview on Nov 11 and launched it Nov 16. What the Qwen4 benchmarks, pricing and closed weights say about the open-weight gap.

Metir AI TeamNovember 21, 20265 min read
Alibaba Qwen4-Max: Benchmarks, Price and Closed Weights

Alibaba's Qwen4 era began quietly. On November 11, 2026, the Qwen team put Qwen4-Max into a managed preview, five days before a full launch, and the most striking thing about the flagship is what Alibaba did not ship with it: downloadable weights.

Alibaba logoAlibaba
Qwen logoQwen
Moonshot AI logoMoonshot AI
Alibaba's Qwen4-Max and Moonshot's Kimi K3.1 Pro are the two models compared in this post.

What Alibaba announced on Nov 11 and Nov 16

The preview opened at 10:00 on November 11 (UTC+8) on Alibaba's Token Plan, with early access also offered through QwenWork, Qoder, the Qwen app and QwenStudio, according to Alibaba Cloud's own product-updates page. The preview model, qwen4-max-preview, was scheduled to retire at 24:00 on November 25. Alibaba's documentation lists a 1-million-token context window, a 393,216-token output ceiling, text, image and video input, and reasoning levels of low, medium and xhigh.

Alibaba's product-updates page records general availability of Qwen4-Max on November 16, and the company's press release of the same date describes it as the most capable Qwen model so far. Preview-era pay-as-you-go access was not supported, so the launch also opened the model to the standard QwenCloud API.

2.4TTotal parametersSparse mixture of experts
71BActive per token25% fewer than Qwen3.8-Max
1MContext windowTokens
$1.50 / $4.50Per 1M tokensInput / output

Architecture, as Alibaba describes it

The press release says Qwen4-Max keeps the 2.4 trillion parameter scale of Qwen 3.8 Max but activates 71 billion parameters per token, 25% fewer, in a sparse mixture-of-experts design with upgrades to attention, residual connections and embeddings. It has native vision capability. Alibaba has not published a technical report in the sources reviewed, so these figures are the company's own.

The white lattice facade and walkways of Alibaba Group's headquarters in Hangzhou
Alibaba Group's headquarters campus in Hangzhou. The photo is illustrative and predates Qwen4. Credit: Thomas LOMBARD, CC BY-SA 3.0.

Qwen4 benchmarks: mixed, and mostly self-reported

Alibaba's launch materials lean on its own tests. It says Qwen4-Max ranks 6th on Code Arena, and that on Work360, an in-house benchmark spanning 721 specialized fields, it beats Qwen3.8-Max while using 35% fewer tokens at high effort and cutting average cost per task by about 60%. The press release carries no independent verification.

Outside measurements are more modest. Decrypt reports a debut at 6th on Code Arena WebDev with 1,750 points, and 14th among individual models on the Arena Agent leaderboard, the highest Chinese entry, while noting that American labs still hold the top overall spots. The Next Web, relaying charts compiled by FourWeekMBA, reports a split result: 75.7 on DeepSWE v1.1 against 74.2 for Claude Opus 5.5, but 48.5 on Terminal-Bench 4.0 against 59.6 for Opus 5.5 and 67.2 for Kimi K3.1.

Artificial Analysis lists Qwen4-Max at 43 on its Intelligence Index, ranked 51st of 95 models, just under the 44 median. It also calls the model very verbose: 127 million output tokens across the index run against a median of 65.4 million, and about 29 minutes per task against a 12.7 minute median.

“

Capability did not move the model to the top of the table. Price per finished task did.

Analysis of the figures reported by Artificial Analysis and Alibaba

Pricing and the cost-per-task story

QwenCloud lists $1.50 per million input tokens and $4.50 per million output tokens, with cached input at $0.075. Decrypt puts that about 25% below Qwen3.8-Max, at $2 and $6. Cellcog, summarizing Artificial Analysis data on November 16, reports an average cost of $3.27 per index task for Qwen4-Max, against $10.02 for Qwen3.8-Max and $3.68 for Kimi K3.1 Pro. Its index scores were 43, 34 and 46.

Qwen4-Max cut the cost of finishing a task by about two thirds

Average USD cost per Artificial Analysis Intelligence Index task, as reported on 16 November 2026. Index scores in brackets.

Cost per task combines per-token price and tokens used, which is why a 25% list-price cut shows up as a much larger drop.

Read plainly, the generation-over-generation gain is real: a nine-point index rise at roughly a third of the cost per task. Against Kimi K3.1 Pro the picture is a trade, three points lower for about 11% less per task, with verbosity doing much of the work in both directions.

Open weights versus the closed frontier

Qwen4-Max is closed. Artificial Analysis lists it as proprietary with no public weights and an undisclosed parameter count, and the press release says only that smaller open-weight Qwen4 models "will soon be released", without dates, sizes or licenses. That is a change from Qwen 3.8 Max, whose weights were opened earlier in 2026, and it comes with Alibaba's own open-source record in the background: over 460 open-weight models and more than 3 billion downloads, per its November 2 announcement as cited by Shattered.io.

The cost-per-task data suggests why a lab might reverse course. If the flagship's selling point is efficiency, the revenue sits in serving it. One report, citing two anonymous people, says Alibaba plans to seek a revenue share from large commercial users of its next open model, though Alibaba has not confirmed it.

What teams should take from this

  • Treat the 60% cost claim as Alibaba's, and the $3.27 per task as the figure to plan against.
  • Expect the open Qwen4 models to be smaller than the flagship, with licensing still unknown.
  • Keep routing flexible. Tools such as Metir, which sit above several model providers, make it easier to switch when a flagship's price, license or openness changes.

The gap between open and closed frontier models did not close with Qwen4. It moved to a different axis, where Alibaba now sells efficiency rather than giving away scale.

Sources:

  • Token Plan product updates | Alibaba Cloud
  • Qwen4-Max-Preview documentation | Alibaba Cloud
  • Alibaba Unveils Qwen4-Max | Alibaba Cloud press room
  • Alibaba launches Qwen4-Max flagship AI model without open weights | The Next Web
  • Alibaba's Qwen 4 Max Leads China's AI Agent Race, But You Can't Download It | Decrypt
  • Qwen4 Max model page | Artificial Analysis
  • Qwen 4 Release Date: Qwen4-Max Is Now in Preview | Cellcog
  • Qwen 4 Launch: Alibaba Open-Source Promise in Doubt | Shattered.io
  • Qwen | Wikipedia

Image credits

Header image: Alibaba Xixi Park campus, Hangzhou, by Danielinblue via Wikimedia Commons, licensed under CC BY-SA 4.0. In-body photograph of the Alibaba Group headquarters in Hangzhou by Thomas LOMBARD via Wikimedia Commons, licensed under CC BY-SA 3.0.

Ready to experience AI that adapts to you?

metir brings together the world's best AI models in one seamless experience. Start for free today.

Get Started Free
metir

metir is where work gets done. Hire your team of Orbs, AI teammates with their own computer, alongside your mail, meetings and workspace.

© 2026 metir. All rights reserved.

Product

  • Features
  • Pricing
  • Research
  • Docs
  • Blog
  • Enterprise

Company

  • Docs
  • Support
  • Careers

Legal

  • Terms of service
  • Privacy policy

Personalisation is powerful. Privacy is non-negotiable.

Status: All systems operational