Qwen4 Max: Nine Points Ahead On Score, Three Times Cheaper, Behind The Pack Leaders
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Qwen4 Max: Nine Points Ahead On Score, Three Times Cheaper, Behind The Pack Leaders on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Alibaba launched Qwen4 Max on Nov. 16, 2026, describing it as a 2.4-trillion-parameter model with vision and a one-million-token context window. Artificial Analysis scored it 43, nine points above the previous Qwen flagship, and measured a lower cost per index task; it ranks ninth among 10 models on its page. The weights are closed, and launch-day speed data was unavailable.

Alibaba launched Qwen4 Max on Nov. 16, presenting it as a 2.4-trillion-parameter model with image and video capabilities and a one-million-token context window. An independent evaluation by Artificial Analysis scored it 43 points, nine above its predecessor, and estimated a lower cost per benchmark task; the model nevertheless ranked ninth among 10 models listed on its comparison page.

Artificial Analysis tested Qwen4 Max at its highest reasoning setting, xhigh, using Intelligence Index v5.0.1, which combines nine evaluations covering areas such as coding, office tasks, knowledge and physics. The firm described 43 as above the median of 41 among models in a similar price range. Qwen3.8 Max (0902) scored 34 on the same index. These results are a comparison of measured performance, not confirmation of Alibaba’s broader launch claims.

Artificial Analysis put the cost of running one index task at $3.27 for Qwen4 Max, compared with $10.02 for Qwen3.8 Max (0902). Alibaba’s listed API prices are $1.50 per million input tokens and $4.50 per million output tokens. The two task-cost figures are specific to Artificial Analysis’s index evaluation; they should not be read as a fixed price for every customer workload.

Alibaba separately reported that its in-house Work360 score rose from 47.74 to 52.69, while average cost per task on that test fell from $4.43 to $1.67. The company also cited improvements in token use and secure-coding results. Those figures come from Alibaba’s own evaluations, so they are not independent findings. Its manufacturing and other customer examples are case studies rather than controlled comparisons.

At a glance
reportWhen: Launched November 16, 2026; independent…
The developmentAlibaba released Qwen4 Max, and an independent benchmark reported a higher score and lower task cost than its predecessor while placing it ninth among 10 models on the comparison page.

Qwen4 Max · launch day

From launch claim to independent check

Alibaba’s numbers come first. An outside measurement shows where the model actually lands.

Alibaba: 2.4T parameters, 71B active Own test Work360: 47.74 → 52.69 Independent index (0 to 70) 0 70 34 41 43 65 Speed: no data yet
  1. The launchAlibaba calls Qwen4 Max a 24/7 “teammate”, with 2.4 trillion parameters and a one-million-token context window.
  2. The vendor’s own testOn its in-house Work360 benchmark, Alibaba reports 52.69, up from 47.74 for Qwen3.8 Max.
  3. The independent indexArtificial Analysis scores it 43, up from 34 and just above the median of 41 for models at a similar price.
  4. What is still openThe top model on the same page scores 65. Speed is not measured yet, and the weights are closed.

Sources: Alibaba/Qwen launch blog, Nov 16, 2026 (vendor figures); Artificial Analysis, model pages for Qwen4 Max and Qwen3.8 Max (0902), retrieved Nov 16, 2026. Schematic scale; marker positions follow the index values.

Graphic: Powered by Thorsten Meyer AI

Lower Benchmark Costs, Modest Ranking

The launch presents a trade-off for teams comparing AI models: Qwen4 Max shows a substantial improvement over the previous Qwen flagship on one independent index, alongside a lower estimated task cost, but it does not sit near the top of the models in the listed comparison. On that page, the leading score is 65, while Qwen4 Max’s 43 places it ninth of 10. A score above the cited price-range median is not the same as leading the overall field.

Cost and capability both matter when organizations choose models for routine work, coding or long-document tasks. But the index task estimate is not a universal measure of an organization’s bill: actual spending depends on usage, token volume, model settings and the task. Alibaba’s claims of lower token use and improvements on particular workloads may be relevant to buyers, but they need to be checked against independent results and each customer’s own tests.

Access is another practical constraint. Qwen4 Max is available through Alibaba services, but its weights are not downloadable. That limits independent inspection and on-premises deployment compared with models whose weights are openly released. It also leaves users reliant on Alibaba’s API and associated services.

As an affiliate, we earn on qualifying purchases.

A Closed First Qwen4 Model

Alibaba called Qwen4 Max the first model in its new Qwen4 family. Its launch blog describes a sparse mixture-of-experts design with 2.4 trillion backbone parameters, 71 billion activated per token and an additional 328 billion N-gram embedding parameters. In this architecture, only part of the network is active for a given token; the headline parameter count alone does not specify the cost or speed of a typical request.

The company says the model uses a mix of Gated DeltaNet layers and “Qwen Sparse Attention” for long inputs, as well as other changes intended to support training and token processing. Alibaba also says training proceeded without loss spikes. Because the weights and training process are not publicly available, outside researchers cannot independently verify those technical descriptions from the model itself.

Alibaba documents text, image and video input, with text output, and says reasoning is enabled by default at the high setting. Artificial Analysis used xhigh reasoning for its index test, so its score may not represent a default API configuration. The model is offered through QwenCloud API, Token Plan and Alibaba apps including QwenWork, Qoder and QwenStudio. A limited preview reportedly began Nov. 11, according to The Next Web.

Alibaba’s launch material includes comparisons with competing models, but its main chart was published as an image. The source report did not independently reconstruct values from that graphic. The Next Web, citing FourWeekMBA, reported mixed results on two coding tests: Qwen4 Max was reported at 75.7 on DeepSWE v1.1, ahead of Claude Opus 5.5 at 74.2, but at 48.5 on Terminal-Bench 4.0, behind Claude Opus 5.5 at 59.6 and Kimi K3.1 Pro at 67.2. These figures are reported comparisons, not Artificial Analysis’s overall index result.

“In our Work360 evaluation, the average cost per task drops from $4.43 to $1.67 — a 62% reduction.”

— Alibaba

Speed, Weights and Real-World Results

Speed and latency data were not available from Artificial Analysis on launch day, so the lower task-cost estimate does not establish that Qwen4 Max responds faster or performs more work per unit time. The benchmark’s score was obtained at xhigh reasoning, while Alibaba’s documented API default is high; the size of any score difference between settings is not given.

The model’s closed weights also limit independent testing of its architecture and training claims. Alibaba says smaller Qwen4 models will be open-sourced but has not announced a date. It is not yet clear when independent evaluations will cover more real-world tasks, how the reported vendor benchmarks compare under matched conditions, or whether customer costs will track the benchmark estimate.

The available evidence supports a narrower conclusion than the launch pitch: Qwen4 Max improved on its predecessor in Artificial Analysis’s index and cost less per task in that evaluation. It does not establish a lead over the strongest models overall. Alibaba’s additional benchmark and customer claims remain attributable to the company unless independently replicated.

Independent Tests and Open-Source Plans

The next useful developments will be additional independent measurements of speed, latency and performance at default settings, alongside tests of the tasks organizations actually plan to run. Comparisons using matched reasoning settings and clearly stated prices would help separate model quality from configuration and workload differences.

Users and developers will also be watching for Alibaba’s promised smaller open-source Qwen4 releases. The company has not given a release date, and the launch does not make Qwen4 Max itself downloadable. Until further data or releases appear, performance claims beyond the published index result and costs should be treated according to their source: independent benchmark figures as measured results, and Alibaba’s Work360 and case-study numbers as company-reported claims.

Key Questions

What did Artificial Analysis score Qwen4 Max?

Artificial Analysis gave Qwen4 Max a score of 43 on Intelligence Index v5.0.1, testing at xhigh reasoning. It reported a median of 41 for comparable models and ranked Qwen4 Max ninth among the 10 models on its page.

How does it compare with the previous Qwen flagship?

Qwen4 Max scored nine points higher than Qwen3.8 Max (0902), which scored 34 on the same index. Artificial Analysis also estimated a lower cost per index task: $3.27 versus $10.02.

Is Qwen4 Max open source?

No. Qwen4 Max’s weights are not available to download; access is through Alibaba services. Alibaba says smaller Qwen4 models will be open-sourced, but it has not announced a date.

Are Alibaba’s Work360 results independently verified?

No independent verification is reported for those figures. Alibaba says its Work360 score rose from 47.74 to 52.69 and its average task cost fell from $4.43 to $1.67. Those should be treated as company-reported results, distinct from Artificial Analysis’s measurements.

Source: ThorstenMeyerAI.com

You May Also Like

The license. Why the AI content market pays the brand-name corpus and strands the long tail.

Large publishers secure licensing deals with AI firms, leaving small publishers excluded and reinforcing market asymmetries. Collective licensing may offer a solution.

No-Code Chrome Extension Builder: An AI-Driven Approach

A new web app enables non-developers to create Chrome extensions using natural language prompts, streamlining browser automation for prosumers and teams.

Musk Fails To Overturn Minnesota’s AI Nudifying Restrictions In Court

A judge denied Elon Musk’s request to temporarily block Minnesota’s law banning AI-generated nonconsensual nude images, allowing enforcement to continue during litigation.

How A Bad Week Helps Prepare AI Agents For Business

Firmulate reports how five AI models handled a simulated company’s crisis week and describes a read-only pilot for business-specific wargames.