Qwen4-Max: What's New, Pricing, and Alternatives (2026)

Alibaba's Qwen team released Qwen4-Max on November 16, 2026. The official post introduces it as "your new AI teammate for work, creativity, and discovery."
It is the first model in the new Qwen4 family. The post describes "a 2.4-trillion-parameter backbone, native vision capabilities, and a one-million-token context window." It also puts heavy weight on finished work, such as spreadsheets and slide decks, not only chat answers.
This guide covers what is new in Qwen4-Max, how it compares with Qwen3.8-Max, where you can use it, and what it costs. It ends with alternatives, including tools that produce reports and slides from your own files. Details come from the Qwen blog and pricing listings, read on November 16, 2026.
What Qwen announced
The release came in two steps. On November 11, IT Home reported that Alibaba had announced Qwen4 and opened a Qwen4-Max preview through its Token Plan. Five days later, the Qwen team published the full release post.
The post sets out the goal in one line. Qwen4-Max "advances real-world productivity through greater proficiency with professional tools, stronger long-horizon task execution, and higher-quality deliverables."
The team lists a wide range of work it can take on. That includes "3D modeling, game development, precision manufacturing, finance, law, healthcare, and frontier scientific discovery."
Qwen4-Max is also only the start of the family. The post says "Max is only the first look at Qwen4." It adds that smaller models in the family "are coming to open source." Qwen4-Max itself is offered through Qwen's platforms, and the post does not mention open weights for it.
What's new in Qwen4-Max
The headline changes fall into four groups.
A new architecture. The post says Qwen4-Max has "2.4T backbone parameters, 71B activated per token, and an additional 328B N-gram Embedding parameters." It is a sparse mixture-of-experts model, so only part of the network runs for each token.
Native vision and long context. The model reads images as well as text. OpenRouter's listing says it accepts "text, image, and video input and returns text." The context window is one million tokens.
Work deliverables. The post's finance example ends with files, not text. It says "deliverables are generated as Excel workbooks and editable PowerPoint decks." Charts in the deck map to the workbook, so analysts can trace each figure back.
Long-running agents. Qwen calls this Live Delivery. The agent "delivers a usable first version early, then expands and improves it through successive releases." Users can step in at any point to change direction. The team is also open-sourcing a Qwen-Live-Delivery-Plugin.
Under the hood: the Qwen4 architecture
The post explains three changes to the model's design. They matter mainly for cost and for long documents.
- Gated DeltaNet plus Qwen Sparse Attention. For every sparse attention layer, three Gated DeltaNet layers compress history into fixed-size states. The sparse layer then selects only the relevant parts of a long context. The post says this reduces "both attention and indexing costs for long sequences."
- Gated Residual. Four parallel residual branches pass information between layers. The team says this improves training stability.
- N-gram Embedding. Lookups based on nearby tokens add capacity "with little additional computation."
The team also reports a smooth training run. It says loss decreased steadily, with "no loss spikes throughout the run."
For users, the practical result is the long context at a lower cost. A one-million-token window can hold a large contract set or a year of reports. The sparse design is how Qwen keeps that affordable.
How it compares with Qwen3.8-Max
Qwen measured the new model mainly against its own previous flagship. The post says Qwen4-Max delivers "stronger performance across nearly every capability we evaluated."
The clearest numbers come from Work360. This is Qwen's own benchmark for professional work. Its taxonomy spans "26 Tier-1 domains, 182 Tier-2 subdomains, and 721 Tier-3 specialized fields."
| Measure | Qwen3.8-Max | Qwen4-Max | Source |
|---|---|---|---|
| Work360 score | 47.74 | 52.69 | Qwen blog |
| Token use on Work360, Xhigh effort | Baseline | About 24% lower | Qwen blog |
| Token use on Work360, High effort | Baseline | About 35% lower | Qwen blog |
| Average cost per Work360 task | $4.43 | $1.67 | Qwen blog |
The cost figure is the most practical one. The post says the average cost per task "drops from $4.43 to $1.67 — a 62% reduction." Part of that comes from using fewer tokens. Part comes from lower API prices, with "particularly large reductions in cached-input pricing."
On coding work, the post reports savings on Terminal-Bench 4.0 too. Token use there drops about 48% and 49% at the two effort levels.
Keep in mind that Work360 is Qwen's own benchmark. Independent tests may show a different picture. The footnotes also cite outside leaderboards for other benchmarks, which helps with comparison.
Example tasks from the release
The post spends most of its length on real tasks. A few are useful for office and analysis work.
Investment research. The post says Qwen4 "invokes 90 MCP tools" to collect market quotes, financials, interest rates, and dividend data. It then builds the analysis. The output is an Excel workbook and an editable PowerPoint deck. Cells link to their sources, and charts match the workbook.
Legal due diligence. In a pre-IPO case, the model received 198 inconsistently named working papers. It renamed and sorted them by content into folders. It then cross-checked them against public records.
Healthcare. The post describes clinical support work that consolidates evidence from several sources for frontline staff.
3D and games. Qwen4-Max uses Blender for modeling and works with game engines such as Unity, Unreal Engine, and Godot.
The pattern across these examples is the same. The model uses tools for many steps and hands back files. That is a shift from earlier releases, which centered on chat and coding.
Where you can use it
Qwen4-Max is available in several places. The post's "How to use" section says it "is now available on QwenCloud." Developers "can start building with either the Token Plan or the API."
The post also names apps where the model appears:
- QwenStudio. The blog links it to the Qwen chat site at chat.qwen.ai.
- QwenWork. The Chinese version of the post names it as Qwen's office suite.
- Qoder. The third product named in the post.
Third-party platforms carry it too. OpenRouter lists the model as qwen/qwen4-max with a one-million-token context window. Its description says reasoning is "enabled by default."
Pricing
Qwen says API prices are lower than Qwen3.8-Max. The release post does not list per-token prices itself, so the table below uses OpenRouter's listing as a reference.
| Item | Price on OpenRouter (per 1M tokens) |
|---|---|
| Input | $1.50 |
| Output | $4.50 |
| Cached input read | $0.075 |
Prices on QwenCloud may differ by region and plan. The Token Plan has its own terms on Qwen's platform. Check the current page before you commit to a budget.
For a sense of scale, compare a typical report task. Suppose a run reads 200,000 tokens and writes 20,000. At the rates above, that is $0.30 for input and $0.09 for output. The total is about $0.39 before any caching.
Caching changes the math for repeated work. If the same documents are read again, cached input costs $0.075 per million tokens. That is one-twentieth of the normal input rate.
What this means for teams
For most teams, the question is not the parameter count. It is whether the model can finish work they would otherwise do by hand.
Qwen4-Max is aimed at that. The finance example produces a workbook and a deck with traceable numbers. The legal example sorts nearly 200 files. Both are tasks that eat up analyst hours.
Three things are worth checking before you adopt it:
- Where you will run it. The API needs a developer setup. QwenStudio is easier but works as a chat app.
- Where your data goes. Read the platform's data terms, especially for client files.
- How you will check results. Long-running agents can drift. Live Delivery helps because you see drafts early.
Start with one recurring task. A monthly market summary or a contract review is a good test. Compare the output and the time spent with your current process.
Limits and open questions
A few points are still unclear or worth caution.
- Self-reported results. The main gains come from Work360, which Qwen built. Wait for outside tests before you rely on the numbers.
- Human review. The post itself says Qwen "is designed to assist people in their work, not to replace professional judgment in high-impact scenarios." Finance, legal, and medical outputs still need an expert check.
- Input data rules. Users must make sure input data "is obtained and used lawfully." This applies to client files and to images of people.
- Regional products. The English and Chinese versions of the post name different app lists. The Chinese version adds Qoder CN and the Qwen App. Check which ones you can use where you are.
None of these is unusual for a new flagship model. They are reasons to pilot before you scale.
Alternatives to Qwen4-Max
Qwen4-Max competes with other frontier models and with tools built on top of them. Each detail below comes from the vendor's own pages.
| Option | What it is | Listed price |
|---|---|---|
| Qwen4-Max | Qwen's new flagship with native vision and a 1M context | $1.50 input / $4.50 output per 1M tokens on OpenRouter |
| Gemini 4 Hydrogen | Google's model for everyday coding and agent work | $1.50 input / $7.50 output per 1M tokens |
| DeepSeek V4.1 Flash | DeepSeek's API model, served as deepseek-flash | $0.3 input (cache miss) / $1.2 output per 1M tokens at peak |
| Powerdrill Bloom | An AI workspace that turns files into reports and slides | Free plan; Pro $13.27/mo |
Gemini 4 Hydrogen. Google's post says it "is priced flat across its entire 1 million-token context window." The rates are $1.50 per million input tokens and $7.50 per million output tokens.
DeepSeek V4.1 Flash. DeepSeek's pricing page lists peak rates for the deepseek-flash model. It notes that off-peak rates are half of the peak rates. We covered the model in our DeepSeek V4.1 Flash guide.
For the previous Qwen generation, see our Qwen3.8 overview. For assistants that compete with Qwen's apps, see our list of Qwen alternatives.
Where Powerdrill Bloom fits
A model is one part of the job. Teams still need a place to bring their files and get a finished report or deck.
Powerdrill Bloom is built for that step. You upload Excel, CSV, or PDF files and describe the output in natural language. The Excel to PPT page says it "turns a topic or your Excel, CSV and PDF files into PowerPoint slides." The AI Report Generator builds reports from the same files.
It works in a web workspace with no code. Files stay in that workspace, so later questions can use them without uploading again.
The pricing page lists a Free plan at $0.00 and Pro at $13.27/mo.
Try Powerdrill Bloom with a quarterly financial export and ask for a summary deck with charts.
Frequently asked questions
What is Qwen4-Max?
Qwen4-Max is the flagship model in Alibaba's Qwen4 family. It has a 2.4-trillion-parameter backbone, native vision, and a one-million-token context window. It is built for long tasks that end in files, such as workbooks and decks.
When was Qwen4-Max released?
A preview opened on November 11, 2026, through Alibaba's Token Plan. The Qwen team published the full release on November 16, 2026.
How much does Qwen4-Max cost?
OpenRouter lists $1.50 per million input tokens and $4.50 per million output tokens. Cached input is $0.075 per million. Prices on QwenCloud may differ.
Is Qwen4-Max open source?
The release post does not mention open weights for Qwen4-Max. It says smaller Qwen4 models are coming to open source.
Where can I try Qwen4-Max?
You can use it in QwenStudio, QwenWork, and Qoder. Developers can access it on QwenCloud through the Token Plan or the API, or through platforms such as OpenRouter.
Sources: Qwen, Qwen4-Max: Your New Teammate Has Arrived · IT Home, Qwen4 preview report · OpenRouter, Qwen4 Max · Google DeepMind, Introducing Gemini 4 Hydrogen · DeepSeek, Models and Pricing.