Qwen 4 Explained
Qwen4-Max is available, while Alibaba has only announced the smaller open-weight models.
Yes, Qwen 4 is out in part. Alibaba Cloud released Qwen4-Max on November 16, 2026. Alibaba calls it the most capable Qwen model to date. Smaller open-weight models in the Qwen4 series have been announced, but Alibaba's November 16 and November 17 posts give no release date, model names, sizes or licence terms for them.
Alibaba's November 16 launch post says Qwen4-Max is available on QwenCloud through the Token Plan and application programming interface (API). It also describes native vision and a context window of one million tokens. The November 17 enterprise post calls Qwen4-Max the flagship of the Qwen4 series. Its benchmark figures compare Qwen4-Max with Qwen3.8-Max, not with models from other providers.
Qwen4-Max Is Released and Smaller Qwen4 Models Are Announced
Qwen4-Max is released, but Alibaba has only announced the smaller open-weight Qwen4 models. Alibaba's November 16 launch post gives November 16, 2026, as the Qwen4-Max release date and says the smaller models will arrive "soon." Neither that post nor Alibaba's November 17 enterprise post names another released Qwen4 model.
Qwen4-Max is the flagship model in the series, according to Alibaba's November 17 enterprise post. It combines native vision with a one-million-token context window, according to both Alibaba posts. The enterprise post uses the name "Qwen4-Max for Enterprise" when discussing business access. That name refers to Qwen4-Max, not a separate model edition.
Anyone who needs to host model weights should note that difference. Alibaba describes future smaller Qwen4 models as open weight, while its November 16 post offers Qwen4-Max through QwenCloud, the Token Plan and the API. The two posts do not describe Qwen4-Max as an open-weight release. For an API evaluation, Qwen4-Max is available. For a Qwen4 open-weight deployment, Alibaba has not published the model or the terms needed to plan one.
What Alibaba Published About Qwen4-Max
Alibaba has published Qwen4-Max's release status, architecture, context limit and access routes. The figures below come from its launch and enterprise posts unless the row names its Model Studio documentation.
| Item | What Alibaba published | Source and date |
|---|---|---|
| Release date | Qwen4-Max released November 16, 2026 | Alibaba launch post, November 16, 2026 |
| Total parameters | 2.4 trillion | Alibaba launch post, November 16, and enterprise post, November 17, 2026 |
| Active parameters per token | 71 billion | Alibaba launch post, November 16, 2026 |
| N-gram embedding parameters | An additional 328 billion | Alibaba enterprise post, November 17, 2026 |
| Architecture | Sparse Mixture-of-Experts (MoE), with Gated DeltaNet, Qwen Sparse Attention, Gated Residual and N-gram embedding | Alibaba launch post, November 16 (sparse MoE upgrade), and enterprise post, November 17, 2026 (named components) |
| Vision | Native vision | Alibaba launch post, November 16, and enterprise post, November 17, 2026 |
| Context window | One million tokens | Alibaba launch post, November 16, and enterprise post, November 17, 2026 |
| Availability | QwenCloud through the Token Plan and API, with access across Alibaba services | Alibaba launch post, November 16, 2026 |
| Live Delivery | An early usable version of a long task, followed by improved versions while work continues | Alibaba enterprise post, November 17, 2026 |
Total and active parameters answer different questions. In a sparse MoE model, a routing step sends each token through a subset of expert sub-networks. Total parameters describe the model's overall size. Active parameters per token are closer to the portion of the model used to process each token. Alibaba's November 16 post reports 2.4 trillion total parameters but 71 billion active parameters per token for Qwen4-Max. It says the active figure is 25% below Qwen3.8-Max at the same total scale.
Alibaba's November 17 enterprise post says Gated DeltaNet layers work with Qwen Sparse Attention to keep very long contexts efficient. The same post describes Live Delivery as a way to review an early usable version and redirect a long task before it finishes. Alibaba says a Qwen-Live-Delivery-Plugin is available alongside Qwen4-Max.
Results Alibaba Reports for Qwen4-Max
Alibaba reports higher Work360 scores and lower token use for Qwen4-Max than for Qwen3.8-Max on its named evaluations. Its November 17 enterprise post describes Qwen Work360 as a framework covering 26 top-level domains, 182 sub-domains and 721 fine-grained domains of professional work. Alibaba's November 16 launch post calls Work360 a proprietary benchmark. The figures below are Alibaba-reported results, not measurements we ran. Only the Code Arena rank compares Qwen4-Max with other providers' models.
Alibaba's November 16 launch post and November 17 enterprise post describe launch demonstrations in manufacturing and science, and the enterprise post adds legal work and cloud operations. The enterprise post labels its results "in the case shown." Demonstration settings differ from other tasks, so the reported times and outputs are examples rather than expected results for a new deployment.
- Alibaba's November 17 enterprise post reports a Work360 score of 52.69 for Qwen4-Max, compared with 47.74 for Qwen3.8-Max. It reports about 24% fewer tokens at Xhigh reasoning effort and about 35% fewer at High, with 58% lower average cost per Work360 task.
- The same post reports about 48% fewer tokens at Xhigh and 49% fewer at High on Terminal-Bench 4.0. These token counts compare Qwen4-Max with Qwen3.8-Max on that evaluation.
- Alibaba's November 16 launch post ranks Qwen4-Max sixth on Code Arena, which the post calls a real-world coding leaderboard. It also reports strong results on Agents' Last Exam, Vals Index, JobBench and Chartography without giving scores for those four evaluations.
- Alibaba's November 17 enterprise post reports that functional and secure code at the function level rose from 22.0% for Qwen3.8-Max to 31.8% for Qwen4-Max on Last Secure Code Bench. It reports that code meeting its security measure at the function, file and repository levels together rose from 8% to 14.7%.
- In the manufacturing case shown, Alibaba's November 16 launch post reports that manual workpiece programming time fell by 95%, from roughly 475 minutes to 24. That figure describes the demonstrated task.
- In the science case shown, Alibaba's November 16 launch post describes a continuous 48-hour agent run through four iterations that found an experimentally validated coolant molecule. Alibaba's November 17 enterprise post adds a 2.7-million-molecule candidate space and a robotic wet lab that validated the molecule as a real sample.
- In the legal case shown, Alibaba's November 17 enterprise post reports that Qwen4-Max sorted 198 initial public offering (IPO) due-diligence files into 14 folders. It delivered an 86-page report covering 12 modules and 14 legal issues, with conclusions cited to source files.
- In the cloud-operations case shown, Alibaba's November 17 enterprise post reports 10.42% lower average model-call scheduling cost for data labeling against a reference baseline. It also reports reaching the optimal offline resource-rescheduling schedule more than 14.61 times faster.
- On cloud scheduling, Alibaba's November 16 launch post gives different figures from the November 17 enterprise post: about 9% lower data-annotation model-call cost and more than 14 times faster offline CPU rescheduling, against 10.42% and 14.61 times. The November 16 post calls its figures benchmark testing, while the enterprise post measures against a reference baseline, so the two should not be read as one measurement.
- On cost per task, Alibaba's November 16 launch post reports a reduction of around 60% against Qwen3.8-Max, while the November 17 enterprise post reports 58% lower average cost per Work360 task.
How to Access Qwen4-Max and What It Costs
Qwen4-Max is available through the QwenCloud Token Plan and API, according to Alibaba's November 16 launch post. Alibaba also names access across its own services. Its November 17 enterprise post says enterprise teams can contact Alibaba Cloud to plan a pilot, while availability, features and pricing vary by region.
The Alibaba Cloud Model Studio qwen4-max documentation, last updated November 17, 2026, lists Singapore prices of $1.50 per million input tokens and $4.50 per million output tokens. For Beijing, Hong Kong, Frankfurt, Virginia and Tokyo, it lists $1.413 per million input tokens and $4.24 per million output tokens. Those are regional token prices, not the cost of a complete task. The Qwen4-Max pricing per token and per task guide covers the pricing decision in more detail.
Alibaba's November 16 launch post separately quotes $0.075 per million input tokens for an implicit cache hit, without naming a region for that figure. It should not be substituted for a regional input list price. Model Studio documentation, last updated November 17, 2026, lists a 1,000,000-token context window. It sets maximum output at 131,072 non-thinking tokens or 262,144 tokens with thinking, and lists low, high and xhigh reasoning effort with high as the default.
The production model ID is qwen4-max. The separate preview ID has different output limits and reasoning defaults. The Qwen4-Max-Preview guide says that preview ID retires after 24:00 UTC+8 on November 25, 2026. Work that needs an ID beyond that retirement should be planned around the production model.
Qwen4-Max Compared With Qwen3.8 Max
Alibaba's November 17 enterprise post reports a higher Work360 score and lower average Work360 task cost for Qwen4-Max than for Qwen3.8-Max. Its November 16 launch post says both models have the same 2.4-trillion-parameter total scale, while Qwen4-Max activates 25% fewer parameters per token. Alibaba's reported benchmark and cost figures do not establish how either model will perform on a different prompt set. The Qwen4-Max vs Qwen3.8 Max upgrade decision provides a focused comparison for that choice.
What Is Still Unpublished
Alibaba has announced smaller open-weight Qwen4 models without publishing the details needed to select or deploy one. Its November 16 launch post says they will be released "soon," and neither that post nor the November 17 enterprise post names another released Qwen4 model.
| Item | Status |
|---|---|
| Qwen4-Max release | Published by Alibaba in its November 16, 2026, launch post |
| Qwen4-Max access | Published by Alibaba in its November 16, 2026, launch post and November 17, 2026, enterprise post |
| Smaller open-weight Qwen4 release date | Not published in Alibaba's November 16 or November 17, 2026, posts |
| Smaller model names and parameter sizes | Not published in Alibaba's November 16 or November 17, 2026, posts |
| Smaller model licence terms | Not published in Alibaba's November 16 or November 17, 2026, posts |
An open-weight release with model names and licence terms would change the answer for a reader who needs to host Qwen 4 weights. A change in Alibaba's published regional prices would change the token-cost comparison. Independent results that disagree with Work360 could also change how much weight to give Alibaba's reported advantage over Qwen3.8-Max. Until those details exist, the announced smaller models cannot be treated as available alternatives to the released Qwen4-Max API.
Who Should Evaluate Qwen4-Max
Qwen4-Max is available for a reader who can evaluate an API model through Alibaba's stated access routes. Alibaba's November 17 enterprise post recommends a scoped pilot on the reader's own workflow to measure impact. Its Work360 cost figure reflects that evaluation's tasks and its comparison with Qwen3.8-Max. It does not supply a cost estimate for a different collection of prompts.
A reader who pays per token can compare the same prompts under the two Qwen models before deciding whether to move work to Qwen4-Max. Record the output needed for each task alongside token use and the applicable regional price. Alibaba's November 17 enterprise post says results vary with the task, tools, data and configuration. The comparison should therefore use the configuration intended for the work, rather than treating the reported 58% Work360 cost reduction as a budget forecast.
Qwen4-Max does not fit a reader who requires self-hosted open weights. That reader should wait for Alibaba to publish the smaller models' weights and licence terms. Work involving regulated legal, financial or engineering decisions needs qualified professional review of model outputs, according to Alibaba's November 17 enterprise post. For an available comparison today, run a small Qwen 4 prompt set through Qwen4-Max and Qwen3.8-Max using the same tasks and review criteria.
Frequently Asked Questions
- Qwen 4 is the series whose flagship model is Qwen4-Max, according to Alibaba's November 17 enterprise post. Alibaba released Qwen4-Max on November 16, 2026. Its November 16 launch post also announces smaller open-weight Qwen4 models for a later release, without giving a date.
- Qwen4-Max is the Qwen4 release named in Alibaba's November 16 launch post and November 17 enterprise post. Alibaba calls it the most capable Qwen model to date in the launch post. The two posts name no other released Qwen4 model.
- Alibaba says in its November 16 launch post that smaller open-weight Qwen4 models will be released soon. Neither that post nor the November 17 enterprise post gives their date, sizes or licence terms. Qwen4-Max itself is offered through Alibaba's API and plans, and the posts do not describe it as open weight.
- No. Qwen4-Max is the production model, while qwen4-max-preview is a separate preview ID with different output limits and reasoning defaults. The Qwen4-Max-Preview guide says the preview ID retires after 24:00 UTC+8 on November 25, 2026.
- Alibaba's November 16 launch post says Qwen4-Max is available on QwenCloud through the Token Plan and API, as well as across Alibaba services. Its November 17 enterprise post says regional availability and features vary. The Model Studio qwen4-max documentation provides the production model's published limits and regional prices.
Plan a Qwen 4 Evaluation
Layer3Labs can help you plan a Qwen4-Max evaluation on your own prompts, with review criteria and regional token costs.
Book a Consultation