Qwen 4 Max: benchmarks, price, and why Alibaba kept its biggest model closed

Alibaba's 2.4-trillion-parameter flagship edges Claude Opus 5.5 on one coding test, trails it on three others, and cuts the cost of a typical task by 62%, by Alibaba's count.

Qwen 4 Max: benchmarks, price, and why Alibaba kept its biggest model closed
+ In short

Alibaba released Qwen 4 Max on 16 November 2026, a 2.4-trillion-parameter AI model with vision and a one-million-token context. Alibaba says it cuts average task cost by 62% versus Qwen3.8-Max and narrowly beats Claude Opus 5.5 on DeepSWE, but it trails on terminal coding, and its weights are not downloadable.

Alibaba's Qwen team released Qwen 4 Max on 16 November, its new flagship AI model and the first of the Qwen4 family. It is a 2.4-trillion-parameter model with built-in vision and a one-million-token context window, and Alibaba says it finishes the average work task for 62% less money than Qwen3.8-Max did. The catch for anyone who loved Qwen for its downloadable models: this one is closed.

Our 24-second version:

Watch on YouTube

What changed with Qwen 4 Max?

Before today, Alibaba's best model was Qwen3.8-Max. Now it is Qwen4-Max, built on a new architecture that the Qwen team's announcement describes in detail:

  • Size: 2.4 trillion "backbone" parameters, of which 71 billion are active for each token, plus 328 billion extra parameters in a new N-gram embedding layer.
  • Inputs: text and images natively, with a context window of up to one million tokens.
  • Cost per task: on Work360, Alibaba's own test of real-world work, the average cost per task fell from $4.43 to $1.67.
  • Fewer tokens: Qwen says it uses about 24–35% fewer tokens than Qwen3.8-Max on Work360 and about half as many on Terminal-Bench 4.0, while scoring higher.
  • Where: QwenCloud (through a Token Plan or the API), plus the QwenWork, Qoder and Qwen Studio apps. A limited preview opened on 11 November, TechNode reports.

Qwen chart: Qwen4-Max scores higher than Qwen3.8-Max with fewer tokens per task on Work360 and Terminal-Bench 4.0 Image: Qwen / Alibaba (Qwen4-Max announcement)

How much does Qwen 4 Max cost?

API access costs $1.50 per million input tokens and $4.50 per million output tokens, according to The Next Web and Crypto Briefing. Alibaba's own post says only that prices are lower than Qwen3.8-Max, with the biggest cut on cached input (text the model has already seen, such as a long system prompt).

Alibaba credits two things for its 62% figure: lower prices and efficiency. A model that needs fewer tokens to finish a job costs less even at the same price per token. That number comes from Alibaba's own benchmark, so treat it as a claim until outside testers repeat it.

Qwen 4 Max benchmarks: is it better than Claude?

On Alibaba's published scores, it depends on the job. Qwen4-Max narrowly leads on DeepSWE v1.1, an agentic coding test: 75.7, against 74.2 for Anthropic's Claude Opus 5.5, 75.2 for OpenAI's GPT-6.1 Sol and 75.5 for Moonshot's Kimi K3.1. It wins clearly on long, multi-step agent tasks such as legal work (HLAB-Vals) and Agents' Last Exam.

It trails elsewhere. On Terminal-Bench 4.0, which tests coding inside a command line, it scores 48.5 against Opus 5.5's 59.6 and Kimi's 67.2. On ProgramBench (rebuilding software from its behaviour) it scores 31.5 against Opus's 65.0, and it is behind on Humanity's Last Exam too.

Chart of Qwen4-Max vs Claude Opus 5.5, GPT-6.1 Sol and Kimi K3.1 on DeepSWE, HLAB-Vals, Terminal-Bench 4.0, ProgramBench and HLE Chart: NextDiff, from scores in the Qwen4-Max announcement

All of these numbers are Alibaba's own. Some are taken from outside leaderboards (Artificial Analysis, Arena, Vals and Mercor, per the post's footnotes), but none of the head-to-head results has been independently re-run yet. For a fuller picture of the rivals, see our look at Kimi K3.1 Pro vs Opus 5.5 and Claude Fable 5.2.

How it works, in plain English

Qwen4-Max is a mixture-of-experts model: instead of running all 2.4 trillion parameters for every word, a router picks a small set of specialist sub-networks, so only 71 billion do the work each time. That is what keeps a giant model affordable to run.

The new pieces are about long inputs. Three "Gated DeltaNet" layers squeeze earlier text into a fixed-size memory, and a new Qwen Sparse Attention layer looks up only the parts of a long document that matter, instead of comparing every token with every other one. Alibaba says this, plus a redesigned residual path, let the 2.4T training run finish with no loss spikes.

What Alibaba says it can do

The announcement leans hard on agents that work for hours. Its examples include a cat-wizard action game it says Qwen4-Max built in Unreal Engine from a written brief, 3D scenes modelled and rendered in Blender, and a 48-hour run in which it designed a new chip-coolant molecule that an automated lab then made and tested.

Cat-mage action game Meow of Magic, which Alibaba says Qwen4-Max built in Unreal Engine Image: Qwen / Alibaba (Qwen4-Max demo)

3D alley scene Alibaba says Qwen4-Max modelled and rendered in Blender Image: Qwen / Alibaba (Qwen4-Max demo)

On safety, Alibaba says it kept training data for defensive security work (finding and fixing flaws) but removed data meant to improve attacks, and added safety training on chemical, biological, radiological and nuclear topics. In its secure-coding test, the share of functions that were both correct and safe rose from 22.0% to 31.8%.

Is Qwen 4 Max open source?

No. Unlike many earlier Qwen releases, Qwen4-Max cannot be downloaded or run on your own hardware; you reach it only through Alibaba's cloud and apps. Alibaba says smaller Qwen4 models "are coming to open source", but it has not named their sizes or a date. If you are unsure what the difference means, our explainer on what open source means for AI covers open weights.

Why it matters

Qwen built its following on downloadable models; The Next Web, citing Bloomberg, calls it the most-downloaded open-weight family. Keeping the flagship closed while open-sourcing smaller siblings is the split OpenAI (gpt-oss) and Google (Gemma) already use, and it suggests Alibaba now sees its top model as a product to sell through QwenCloud, not a calling card.

For teams, the pitch is cost: near-frontier agentic coding at a lower price per finished task. The caveats are that the benchmarks are self-reported, it trails Claude and Kimi on terminal-style coding, and closed weights mean no self-hosting. Expect independent results from Artificial Analysis and Arena within days, and watch for the open Qwen4 models.

Questions readers ask

Is Qwen 4 Max open source?
No. Qwen4-Max is available only through Alibaba's QwenCloud and its apps; the weights cannot be downloaded. Alibaba says smaller Qwen4 models are coming to open source but has not given sizes or a date.
How much does Qwen 4 Max cost?
API access costs $1.50 per million input tokens and $4.50 per million output tokens, according to The Next Web. Alibaba says the average task on its Work360 benchmark now costs $1.67, down from $4.43 with Qwen3.8-Max.
Is Qwen 4 Max better than Claude Opus 5.5?
On Alibaba's own published scores, it leads narrowly on DeepSWE (75.7 vs 74.2) and clearly on some long agent tasks, but trails Opus 5.5 on Terminal-Bench 4.0 (48.5 vs 59.6), ProgramBench and Humanity's Last Exam. No independent re-test has been published yet.
How big is Qwen 4 Max?
Alibaba says it has 2.4 trillion backbone parameters, of which 71 billion are active per token, plus 328 billion N-gram embedding parameters, and a context window of up to one million tokens.
Where can I use Qwen 4 Max?
It is available on QwenCloud through a Token Plan or the API, and in Alibaba's QwenWork, Qoder and Qwen Studio apps. A limited preview opened on 11 November and the model launched on 16 November 2026.

+ Sources

  1. Qwen team — Qwen4-Max: Your New Teammate Has Arrived
  2. The Next Web — Alibaba launches Qwen4-Max flagship AI model without open weights
  3. TechNode — Alibaba launches Qwen4-Max with lower API prices and a million-token context window
  4. South China Morning Post — Alibaba says new flagship AI model cuts task costs 60% while tightening cyber guardrails
  5. Crypto Briefing — Alibaba unveils Qwen4-Max, claims it beats rival models on coding

+ Rate this story

How useful was this? Tap a star.
+ The NextDiff app · free

Get stories like this on your phone, the minute they drop.

Get it on Google PlayQR code to install the NextDiff app from Google PlayScan with your Android phone
Android now on Google Play · iPhone coming soon · More about the app →
+ Enjoyed this?Get The Weekly Diff — the 5-minute recap of everything that changed in tech.
+ Watch & discuss this storyThis story in 60 seconds, and the conversation around it.

Keep reading