What Is Qwen4-Max? The New Qwen4 Flagship With 2.4T Backbone and 1M Context (2026)

Ethan Coleon 3 hours ago

What Is Qwen4-Max? The New Qwen4 Flagship With 2.4T Backbone and 1M Context (2026)

Short answer: Qwen4-Max is Alibaba’s official Qwen4 flagship — a hosted multimodal model for coding, office work, frontend design, and long-horizon agent tasks. The official Qwen blog describes a new Qwen4 architecture with a 2.4-trillion-parameter backbone, 71 billion parameters activated per token, an additional 328 billion N-gram Embedding parameters, native vision, and a 1-million-token context window. It is not the October 2026 Qwen4-27B leak rumor, and it is not an open-weight download today.

This page is a plain-English explainer of the official launch. Specs below come from the Qwen blog and Alibaba Cloud Model Studio docs, not from unofficial mirrors or invented benchmarks.

TL;DR

FactOfficial detail
What it isNext-generation Qwen flagship for text generation, image and video understanding, and deep reasoning
Architecture (Qwen blog)New Qwen4 design: 2.4T backbone, 71B activated per token, 328B N-gram Embedding
Context1M tokens (Model Studio and the official blog)
VisionNative image and video understanding; Model Studio says these models respond with text — this is not the Qwen Image generator line
Designed forCoding, office productivity, frontend design, and complex agentic work
ToolsFunction calling, web search, built-in tools, structured JSON output
API styleCompatible with OpenAI and Anthropic APIs
Previewqwen4-max-preview from 10:00 on November 11, 2026 to 24:00 on November 25, 2026 (UTC+8), Token Plan only
General / international listingqwen4-max appears on Alibaba Cloud Model Studio’s release list for November 16, 2026
Singapore list price (Model Studio, updated Nov 16, 2026)$1.50 / $4.50 per million input / output tokens (0 to 1M tokens in a request)
Open weights?The official blog says smaller Qwen4 models are coming to open source. It does not release Qwen4-Max weights in that post
Related NSFW questionSee Can Qwen4-Max do NSFW? — we do not invent test results

What is Qwen4-Max?

Qwen4-Max is the first public Qwen4-series flagship from the Qwen team. The official post, Qwen4-Max: Your New Teammate Has Arrived (cited November 2026), frames it as a 24/7 “AI teammate” for work, creativity, and discovery: stronger professional tool use, longer-horizon task execution, and higher-quality deliverables than the previous Max generation.

Alibaba Cloud Model Studio uses almost the same product blurb: a flagship multimodal model that combines text generation, image and video understanding, and deep reasoning, aimed at coding, office productivity, frontend design, and complex agentic tasks.

In plain English, people searching what is qwen4-max or qwen 4 max are usually asking three things:

  1. Is this a real, official model — or another leak?
  2. How big is it, and what can it see?
  3. Where can I actually call it?

The short version: it is official and hosted. The architecture numbers below are from the Qwen blog, and the API limits are from Model Studio.

Official timeline (don’t mix the dates)

Date (UTC+8 unless noted)What official docs say
November 11, 2026, 10:00qwen4-max-preview preview starts; preview model card says access is Token Plan only, not pay-as-you-go
November 16, 2026Model Studio release notes list qwen4-max for the International deployment scope
November 25, 2026, 24:00Preview period ends; the preview card says the preview model will be retired after that

The official blog itself is the architecture write-up. Its citation block dates the article November 2026. Some HTML metadata inside the page still shows a leftover “August 3, 2026” stamp — treat that as a template leftover, not the launch day.

A short Token Plan / Qwen app preview also appeared around November 11 on Qwen’s public channels. That early access is the preview ID, not a separate secret model.

Architecture: 2.4T backbone, 71B activated, 328B N-gram Embedding

These figures are from the official Qwen blog, not from third-party rumors:

  • 2.4T backbone parameters
  • 71B activated per token (sparse MoE-style activation, not a dense 2.4T forward pass)
  • 328B additional N-gram Embedding parameters
  • Upgrades across attention, residual connections, and embeddings

The blog names three design pieces in plain language:

PieceWhat the Qwen team says it does
GDN + Qwen Sparse Attention (QSA)Gated DeltaNet layers compress history into fixed-size states; QSA uses a lightweight indexer to pick relevant context in small blocks, which is meant to cut long-sequence attention cost
Gated Residual (GR)Parallel residual branches with dynamic read/write gates, aimed at information flow and training stability
N-gram EmbeddingContext-based lookups that add local token-pattern representations with relatively little extra compute

The post also claims pre-training at this scale was stable (steadily falling loss, no loss spikes in their write-up) and that the Qwen4-Max-Base checkpoint improved in-house bits-per-byte versus Qwen3.8-Max-Base by about 5–13% while activating about 25% fewer parameters.

Don’t confuse this with Qwen3.8-Max. The previous Max generation was also described as a 2.4T class model, but official Qwen3.8 materials use a ~95B activated figure (for example qwen3.8-2.4t-a95b). Qwen4-Max’s official activation number is 71B per token, plus the separate N-gram embedding pool.

Context, vision, and API limits

ParameterOfficial sourceValue
Context lengthBlog + Model Studio1M tokens
Input typesModel StudioText, images, and video
Output type (understanding models)Text generation and visual understandingText — you do not switch to a separate VL model ID just to send an image or video
Image / video caps (qwen4-max, International table, updated Nov 16, 2026)Same docUp to 16M pixels per image, 2-hour video, 2 GB video, 2048 images, 64 videos
Max output (qwen4-max, same International table)Same doc131,072 tokens (non-thinking); 262,144 tokens (thinking + output)
Max output (qwen4-max-preview card)Preview model card393,216 tokens, shared by thinking and the final answer
ThinkingPreview cardEnabled by default; intensity is adjusted with reasoning_effort

Those max-output numbers are not identical across the preview card, the International feature table, and some catalog pages. If you are wiring a client, use the doc for your model ID and region rather than flattening them into one invented limit.

How Qwen4-Max compares with nearby Qwen names

NameWhat it actually is
Qwen4-Max (this page)Official Qwen4 flagship, hosted, 2.4T backbone / 71B activated, 1M context, native vision understanding
Qwen4-Max-PreviewThe Token Plan preview ID (qwen4-max-preview), time-boxed through November 25, 2026 (UTC+8)
Qwen4-27B leak chatterAn unconfirmed October 2026 dense-27B rumor. Not this launch
Qwen3.8-MaxPrevious Max generation (also a 2.4T-class MoE, different activation count)
Qwen ImageA separate image-generation line. See Can Qwen Image 2.1 generate NSFW?
Qwen3.8 Omni FlashAn earlier omni / realtime family member, not Qwen4-Max

Where to use it, and what it costs

The official blog says Qwen4-Max is available on QwenCloud, through either the Token Plan or the API, and also in QwenWork, Qoder, and QwenStudio.

Alibaba Cloud Model Studio’s pricing page (updated November 16, 2026) lists Singapore / International qwen4-max at $1.50 per million input tokens and $4.50 per million output tokens for requests whose input is up to 1M tokens, in both thinking and non-thinking modes. A 1 million token free quota is noted for Singapore on that table. Cache prices on the same page include $1.875 / $3.00 per million tokens for 5-minute / 1-hour explicit cache writes and $0.075 per million for cache reads.

The Qwen blog also says list prices were cut versus Qwen3.8-Max, with a large drop in cached-input pricing, and quotes an internal Work360 average cost per task falling from $4.43 to $1.67 (about 62%). Treat that as the vendor’s evaluation cost, not your invoice.

The same blog reports token-usage drops versus Qwen3.8-Max of about 24% and 35% on Work360 at Xhigh and High effort, and about 48% and 49% on Terminal-Bench 4.0. Those are official self-reported efficiency claims, not an independent audit.

What Qwen4-Max is — and is not — good for

A good fit, according to official materials:

  • Long-context coding and agent workflows (tool calling, web search, structured output)
  • Office and professional document work, including visual inputs such as screenshots, charts, and long videos
  • Frontend and software-delivery tasks the Qwen team highlights in the launch post
  • Hosted use through QwenCloud / Model Studio if you do not want to run a 2.4T-class model locally

Not what this launch is:

  • A public open-weight Qwen4-Max checkpoint (smaller Qwen4 models are only promised)
  • A drop-in Qwen4-27B local GGUF (that name is a different, unconfirmed rumor)
  • An image or video generator. Native vision here means the model can read images and video and answer in text
  • An uncensored adult model. If that is your actual question, read Can Qwen4-Max do NSFW?

If you mainly need a browser workspace for creative drafting while you evaluate official Qwen endpoints, HackAIGC Chat is a hosted option. It does not claim to serve official Qwen4-Max weights.

FAQ

Is Qwen4-Max officially released?

Yes. The Qwen team published an official model blog at qwen.ai/blog?id=qwen4-max, a preview ID shipped on November 11, 2026, and Model Studio lists qwen4-max on November 16, 2026.

What does 2.4T / 71B / 328B mean?

Per the official blog: a 2.4-trillion-parameter backbone, 71 billion parameters activated per token, and 328 billion extra N-gram Embedding parameters. It is a sparse MoE-style design, not a dense 2.4T model that activates every parameter on every token.

Does Qwen4-Max have a 1M context window?

Yes. Both the official blog and Alibaba Cloud Model Studio document a 1-million-token context length.

Can Qwen4-Max see images and video?

Yes — official docs describe native image and video understanding. Model Studio’s understanding models reply in text. That is different from Qwen’s image-generation models.

Is Qwen4-Max the same as the Qwen4-27B leak?

No. The Qwen4-27B page covers an unconfirmed October 2026 dense-27B rumor. Qwen4-Max is the official hosted flagship.

Are the weights open source?

Not in the official Max launch post. The team says smaller Qwen4 models are coming to open source. Wait for an official model card before trusting random “Qwen4” uploads.

How do I call the API?

Use the model ID for your surface — qwen4-max on the general International listing, or qwen4-max-preview on the Token Plan during the preview window — and follow the current Model Studio or QwenCloud endpoint docs. The APIs are documented as compatible with OpenAI and Anthropic formats.

Bottom line

Qwen4-Max is the official Qwen4 flagship: a 2.4T-backbone, 71B-activated, native-vision, 1M-context hosted model for coding, office work, and agents. It is a real product with public docs, not a leak. If you want the adult-content angle, use the NSFW companion guide. If you want a hosted creative chat workspace that is not a Qwen official endpoint, start at chat.hackaigc.com.


Sources: Qwen official blog: Qwen4-Max; Model Studio release notes (updated Nov 16, 2026); qwen4-max-preview model card; Text generation and visual understanding; Model Studio pricing.

Related reading: