Sulat.com / Writing on AIMonday, November 16, 2026
Artificial Intelligence/16 November 2026

Alibaba launches Qwen4-Max, a 2.4T-parameter model at US$1.50 per million input tokens

Alibaba's Qwen team launched Qwen4-Max, a 2.4-trillion-parameter model with a 1M-token context, at US$1.50 per million input tokens. Qwen says it cuts the cost of a task by 62%.

Glossy blue 3D Qwen logo feeding a data stream into a glass AI brain ringed by 3D, game and spreadsheet icons

At a glance

  • Alibaba's Qwen team launched Qwen4-Max on 16 November, a model with a 2.4-trillion-parameter backbone, native vision and a one-million-token context window, available through the QwenCloud API.
  • Qwen4-Max costs US$1.50 per million input tokens and US$4.50 per million output tokens, with implicit-cache input at US$0.075. Qwen says every price is below Qwen3.8-Max's.
  • Qwen says it beats Qwen3.8-Max with fewer tokens: about 24% to 35% fewer on Work360 and 48% to 49% fewer on Terminal-Bench 4.0.
  • In Qwen's own Work360 evaluation, the average cost per task falls from US$4.43 to US$1.67. Qwen's post on X calls it an internal evaluation.
  • Qwen says smaller Qwen4 models are coming as open weights. It hasn't given sizes or release dates.

Alibaba’s Qwen team has launched Qwen4-Max, a model with a 2.4-trillion-parameter backbone, native vision and a one-million-token context window, Qwen says in its launch blog. The API is live on QwenCloud at US$1.50 per million input tokens and US$4.50 per million output tokens.

Qwen announced it on X on 16 November. Alibaba opened a preview of Qwen4-Max on 11 November, but only on its Token Plan subscription, with no benchmarks and no per-token price. The launch adds both.

What’s inside

Qwen4-Max is a mixture-of-experts model. Only part of it runs for each token: 71 billion of the 2.4 trillion backbone parameters, Qwen says. Another 328 billion sit in N-gram embeddings, which are lookups for short runs of tokens.

Qwen4-Max spec card: 2.4T parameters, 71B active, 1M context, 393K output, US$1.50 in and US$4.50 out
A 2.4 trillion-parameter model that wakes up 71 billion of them for each token.

The attention layers are new. Qwen says it pairs three Gated DeltaNet layers, which compress history into fixed-size states, with every Qwen Sparse Attention layer. The sparse layer uses a small indexer to pick which parts of a long prompt to read. Four parallel residual branches with read and write gates carry information between layers.

Pre-training ran without a loss spike, Qwen says. On its in-house base-model tests, Qwen4-Max-Base scores 5% to 13% lower bits per byte than Qwen3.8-Max-Base, with 25% fewer active parameters. Bits per byte measures how well a model predicts text. Lower is better.

Fewer tokens, lower cost

Qwen says Qwen4-Max beats Qwen3.8-Max while using fewer tokens. On Work360, token use drops about 24% at xhigh effort and 35% at high. On Terminal-Bench 4.0 it drops about 48% and 49%.

Token savings versus Qwen3.8-Max: Work360 24% and 35%, Terminal-Bench 4.0 48% and 49%, Alibaba says
The savings are biggest on Terminal-Bench 4.0, where Alibaba says Qwen4-Max uses about half the tokens.

Work360 is Qwen’s own benchmark, covering 26 top-level domains, 182 subdomains and 721 specialised fields. Qwen4-Max scored 52.69 on it. Qwen3.8-Max scored 47.74. In that evaluation the average cost per task falls from US$4.43 to US$1.67, a 62% cut. Qwen’s X post calls it an internal evaluation. The blog’s comparison chart also cites results from Mercor, Artificial Analysis, LM Arena and Vals.

Work360 cost per task: Qwen3.8-Max US$4.43 versus Qwen4-Max US$1.67, a 62% drop, Alibaba says
Alibaba says Qwen4-Max cuts the average cost of a Work360 task from US$4.43 to US$1.67.

Price and access

Qwen says every price is lower than Qwen3.8-Max’s, and the implicit-cache price is down 70%. Input can be text, images or video, and output is text. QwenCloud lists a 991,000-token input cap, 393,000 tokens of output and 262,000 for reasoning. It also lists function calling, structured outputs, web search and caching. Qwen4-Max also runs in the Token Plan, QwenWork, Qoder and QwenStudio.

ItemUS$ per million tokens
Input1.50
Output4.50
Input, implicit cache0.075
Explicit cache creation, 5 minutes1.875
Explicit cache creation, 1 hour3.00
Explicit cache read0.075
Qwen4-Max API prices, from QwenCloud’s model page.

What Qwen showed it doing

In game development, Qwen says Qwen4-Max built playable games in Unreal Engine, Unity and Godot from a written brief. It also drove Blender to model and render 3D scenes. At SANF and Pacific Union, programming a single workpiece fell from about 475 minutes of manual effort to 24. For a finance task it called 90 MCP tools. In a legal due-diligence case it sorted 198 working papers into 14 folders and wrote an 86-page report.

The science demos are longer. On Deep Principle’s Mira platform, Qwen says the model ran for 48 hours, screened 2.7 million candidate molecules and directed an automated lab to make a physical sample of a new coolant. On DP Technology’s PharmMaster platform it made 454 tool calls and designed 24 candidate molecules for a protein linked to Alzheimer’s. Qwen puts the top pick’s nominal binding affinity at about 3.6 times the reference molecule’s. DP Technology’s experts think it’s worth wet-lab testing, Qwen says. Those runs happened on partner platforms, and the results are Qwen’s reports.

What’s still to come

Qwen is open-sourcing a Live Delivery plugin. Its job is to have a long-running agent ship a usable first version early, then keep improving it while the user steers. On Last Secure Code Bench, Qwen says the share of Qwen4-Max code that passes both functionality and security checks at function level rose from 22.0% to 31.8%.

Qwen says smaller Qwen4 models are coming as open weights. It hasn’t given sizes or dates.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.