i Informopedia.com
Artificial Intelligence Tier 1 · Public

What Is Qwen4-Max? Alibaba’s New Flagship AI Model, Explained

1,987 words 10 min read Updated November 17, 2026

Qwen4-Max is Alibaba’s new flagship AI model — the first release in the Qwen4 generation, launched on November 16, 2026. It is a 2.4-trillion-parameter model with native vision and a one-million-token context window, built for long professional jobs rather than quick chat: 3D modelling, game development, financial research, legal due diligence and scientific discovery. It costs $1.50 per million input tokens through Alibaba’s QwenCloud service, and Alibaba says it completes industry tasks for 62% less than its predecessor. The catch is that the first independent tests rank it well behind Claude and GPT-6 on overall intelligence — so the real question is not whether Qwen4-Max is the best model, but whether it is good enough at a third of the price.

This guide explains what Qwen4-Max is, how it works, what it costs, how to use it, how it compares with its rivals, and where it falls short. Benchmark figures from Alibaba are labelled as the company’s own; independent figures come from Artificial Analysis, which tests every model through the same harness.

What is Qwen4-Max?

Qwen4-Max is the flagship of Alibaba’s Qwen family, the successor to Qwen3.8-Max and the first model built on a new Qwen4 architecture. Alibaba calls it a “24/7 AI teammate for work, creativity, and discovery” — a model designed less for answering single questions and more for carrying multi-step projects to a finished deliverable: a rendered 3D scene, a playable game, an 86-page legal report, a drug candidate.

It is a sparse mixture-of-experts (MoE) model. Instead of activating its whole network for every token, it routes each token through a subset of experts, which keeps the compute bill down despite the size. The headline numbers:

  • 2.4 trillion backbone parameters, with 71 billion activated per token — the same total size as Qwen3.8-Max, but with 25% fewer active parameters per token.
  • 328 billion N-gram embedding parameters, a new layer that looks up local token patterns to expand capacity cheaply.
  • Native vision: it processes images alongside text rather than bolting on a separate vision encoder.
  • One-million-token context window — roughly a thousand pages of text, or a whole codebase, in one conversation. Independent tracking lists the usable window slightly lower, at 984,000 tokens.
  • New architecture: Gated DeltaNet layers plus Qwen Sparse Attention for long sequences, and four parallel gated residual branches for training stability. Alibaba says pre-training at this scale ran with no loss spikes.

On Alibaba’s in-house tests of the base model across knowledge, code, question answering and STEM, Qwen4-Max-Base improved on Qwen3.8-Max-Base by about 5–13% while activating a quarter fewer parameters. Those are the company’s own numbers, and some rival checkpoints in its charts were post-trained models, which can skew comparisons.

What changed from Qwen3.8-Max

Qwen3.8-Max Qwen4-Max
Backbone parameters 2.4 trillion 2.4 trillion
Active per token ~95 billion 71 billion
Context window 1 million tokens 1 million tokens
Vision Native vision Native vision
Work360 score (Alibaba’s benchmark) 47.74 52.69
Average cost per Work360 task $4.43 $1.67 (−62%)
API input price $2.00 per million tokens $1.50 per million tokens
API output price $6.00 per million tokens $4.50 per million tokens
Artificial Analysis Intelligence Index 34 43
Weights Open-weight Closed — cloud only

The table’s last row is the biggest change. Qwen built its reputation on open weights — the Qwen family is the world’s most-downloaded open-weight model line — and Qwen4-Max breaks that pattern. It is available only through Alibaba’s cloud. Smaller Qwen4 models are promised as open source later, with no date given.

Qwen4-Max price and how to access it

The Qwen4-Max price on QwenCloud is $1.50 per million input tokens and $4.50 per million output tokens, with implicitly cached input at $0.075 per million — about 25% less per token than Qwen3.8-Max, and a 70% cut on cached input. Those list prices come from Alibaba Cloud’s model page; the launch blog itself does not print them.

There are two ways to get access:

  • QwenCloud API (pay-as-you-go). Create an account at QwenCloud, enable the Model Studio API, and call the model with the ID qwen4-max. This is the production model.
  • QwenCloud Token Plan. A subscription route for lighter or predictable use. Note that the qwen4-max-preview model, which opened on November 11 on Token Plan only, is a separate, time-boxed preview that ends on November 25, 2026 and will then be retired — use the production model instead.

The model is also available inside Alibaba’s own products: QwenWork, Qoder (its coding agent) and QwenStudio (chat.qwen.ai). Third-party coding tools such as Claude Code and OpenCode can also be pointed at the QwenCloud API.

What Qwen4-Max can do

Alibaba’s announcement is built around long-horizon agent work — tasks that run for hours and end with a usable deliverable. The showcases are the company’s own, and none has been independently reproduced, but they show what the model is aimed at:

  • 3D modelling and games. In Blender it builds geometry, scenes, materials, lighting and renders; with Unity, Unreal Engine and Godot it takes a natural-language brief to a playable game, testing and refining through visual feedback. One demo built a cat-mage adventure in Unreal Engine.
  • Finance, law and healthcare. For investment research it calls 90 tools to pull market data across Chinese, Hong Kong and US equities, then produces Excel workbooks and PowerPoint decks with source links in the cells. For legal due diligence it sorted 198 inconsistently named working papers into 14 folders, cross-checked public records, and delivered an 86-page report covering 14 legal issues. For a critical-care case it reviewed six days of notes, lab reports and 1,433 medical images and proposed management steps matching the attending physician’s orders.
  • Long-running agents. A feature called Live Delivery hands back a usable first version early and improves it in stages, so users can correct mistakes mid-task instead of waiting for the end. Alibaba also open-sourced the Qwen-Live-Delivery-Plugin for it.
  • Science. In a 48-hour autonomous run on a materials platform, the model screened 2.7 million candidate molecules and directed an automated wet lab to synthesise and test a new coolant for data-centre chips, ending with a physical sample. In drug discovery it made 454 tool calls and designed 24 candidate molecules against an Alzheimer’s-related target.
  • Cloud operations. It schedules Alibaba Cloud’s own GPU and CPU resources, cutting annotation-scheduling costs by about 10% in the company’s tests.

For background on how agents like these work, see our guide to agentic AI.

How good is it really? Qwen4-Max vs Claude and GPT-6

This is where the launch story splits in two. Alibaba’s own comparison chart shows Qwen4-Max beating Claude Opus 5.5 and GPT-6 Astra on DeepSWE, an agentic coding test (75.7 vs 74.2 and 74.1), and far ahead on HLAB-Vals, a legal agent test (21.3 vs 9.2 and 7.1). But on the same chart it trails badly on Terminal-Bench 4.0 (48.5 vs 59.6 for Opus 5.5 and 67.2 for Kimi K3.1 Pro) and on ProgramBench (31.5 vs 65.0).

The independent picture is cooler still. Artificial Analysis scores Qwen4-Max at 43 on its Intelligence Index, ranking 38th of 81 models and just above the median of 41. On the same index: Claude Fable 5.2 scores 65, Gemini 4 Carbon 60, Claude Opus 5.5 59, GPT-6 Astra 57 and Kimi K3.1 Pro 46. Qwen4-Max is ninth among the ten leading models the service compares directly, and third among Chinese models behind Kimi. It is a real step up from Qwen3.8-Max, which scored 34 — but it is not a frontier leader.

Where it wins is cost. An Intelligence Index task costs $3.27 on Qwen4-Max, against $10.04 for Claude Opus 5.5 and $26.54 for Claude Fable 5.2; only GPT-6.1 Sol is cheaper, at $1.27. On LM Arena’s web-development leaderboard it ranks sixth at 1,750 Elo, behind Claude Fable 5.2, Opus 5.5, Gemini 4 Carbon and GPT-6 Astra — competitive, not leading. For comparison with the other November flagships, see our guides to Gemini 4 Hydrogen, Claude Fable 5.2 and GPT-6 Cyber.

The fair summary: Qwen4-Max is a capable, cheap workhorse for long professional and agent tasks, strongest in legal and knowledge work, weakest in terminal coding. If your job is measured in cost per finished task rather than top benchmark scores, the price gap matters more than the nine-point gap to Kimi or the sixteen-point gap to Opus 5.5.

Limits and open questions

  • Closed weights. You cannot download Qwen4-Max, inspect it, or run it on your own hardware. That also means outsiders cannot verify Alibaba’s safety and benchmark claims by testing the model directly.
  • Vendor benchmarks. Work360, the benchmark behind the 62% cost-cut headline, is Alibaba’s own test of its own model. The independent index confirms the direction — cheaper per task, better than its predecessor — but not the size of the gain.
  • Verbosity. Artificial Analysis notes the model generated 127 million output tokens to complete its index, well above the peer median of about 79 million. Long outputs cost money at $4.50 per million output tokens, and they take time: output speed is listed at 53 tokens per second, against 110–117 for the Claude models.
  • No external safety audit yet. Alibaba reports stronger secure-coding results and no observed large-scale cybersecurity risk, but all assessments are internal. Reuters noted the findings had not been independently verified.
  • Preview confusion. The Token Plan preview and the production model share a name but are different models with different end dates. Anyone who started on the preview needs to move to qwen4-max before November 25.

Safety: what Alibaba says

Alibaba says it added code-security checks on top of functionality checks in training: the share of generated code that is both functional and secure rose from 22.0% with Qwen3.8-Max to 31.8% at function level, and from 8% to 14.7% when all three granularities — function, file and repository — must pass at once. For cybersecurity, it says it kept data for defensive tasks such as vulnerability detection and remediation while removing data for offensive attack tasks, and that internal reviewers of full task trajectories found substantially improved defence and no large-scale risk. For chemical, biological, radiological and nuclear topics, it added dedicated safety data in post-training. All of this is Alibaba’s own reporting; treat it as the company’s claims until an external evaluation appears.

Who should use Qwen4-Max?

Consider it if you run long agent workflows — document-heavy research, legal review, data analysis, 3D or game prototyping — and your budget is the binding constraint. At $1.50 per million input tokens with cheap caching, it is one of the lowest-priced flagship-class APIs available, and its legal-agent and long-horizon results are genuinely strong.

Look elsewhere if you need the best coding agent available: Claude Opus 5.5 and GPT-6 Astra both beat it clearly on terminal coding, and Claude Fable 5.2 leads the independent index by 22 points. Skip it too if you need open weights, on-premises deployment, or a model you can fine-tune yourself — Qwen4-Max offers none of those, and the promised smaller open Qwen4 models have no release date.

Frequently asked questions

What is Qwen4-Max?

Qwen4-Max is Alibaba’s flagship AI model, launched November 16, 2026 — the first in the new Qwen4 family. It is a 2.4-trillion-parameter mixture-of-experts model with native vision and a one-million-token context window, designed for long professional and agent tasks.

How much does Qwen4-Max cost?

$1.50 per million input tokens and $4.50 per million output tokens through the QwenCloud API, with implicitly cached input at $0.075 per million. Alibaba says the average cost per industry task fell 62% from its predecessor, from $4.43 to $1.67 on its own Work360 benchmark.

Is Qwen4-Max open source?

No. Unlike earlier Qwen flagships, Qwen4-Max is closed-weight and available only through Alibaba’s cloud services. Alibaba says smaller Qwen4 models will be released as open source later, but has given no date.

How do I use Qwen4-Max?

Through the QwenCloud API with model ID qwen4-max, via a QwenCloud Token Plan, or inside QwenWork, Qoder and QwenStudio. The qwen4-max-preview model is a separate preview that ends November 25, 2026.

Is Qwen4-Max better than Claude or GPT-6?

Not overall. Independent testing by Artificial Analysis scores it 43, against 59 for Claude Opus 5.5 and 57 for GPT-6 Astra. It leads on some tests — DeepSWE coding and a legal agent benchmark — but trails on terminal coding, and it is much cheaper per task than either rival.

What is Qwen4-Max best at?

Long professional work: legal due diligence, financial research, 3D modelling, game development and multi-hour agent runs. Its 1M-token context and Live Delivery feature, which returns usable results in stages, suit jobs that take hours rather than seconds.

More in Artificial Intelligence

Artificial Intelligence Tier 1 · Public

How to Use Claude Meetings: Claude in Google Meet, Explained

Claude Meetings puts Anthropic’s AI assistant inside your video call as a visible participant — not as a recording bot that sends a summary afterwards, but as a colleague that follows the…

2,466 words 12 min read November 14, 2026
Artificial Intelligence Tier 1 · Public

How to Use ChatGPT Wallet: OpenAI’s Agent Payments, Explained

ChatGPT Wallet is OpenAI’s way of letting its AI agents pay for things on your behalf — without the agent, or the shop, ever seeing your real card number. Announced on November…

2,449 words 12 min read November 14, 2026