Citation and evidence

Qwen4-Max

Report a problem with this article

Edit article
More

Cite this article

Qwen4-Max is a proprietary multimodal model released by Alibaba Cloud on November 16, 2026. It is the first production model in the Qwen family's fourth generation, following a separate preview opened on November 11. It accepts text, images and video and produces text, combining native visual understanding with reasoning and tool use. Alibaba offers it through QwenCloud's API and Token Plan, as well as QwenWork and QwenStudio.[1][2][3][4]

Alibaba describes its architecture as a sparse mixture of experts with a 2.4-trillion-parameter backbone and 71 billion parameters activated per token. Its technical launch post separately lists an additional 328 billion N-gram Embedding parameters. Those categories should be kept separate from the press release's shorter description of a 2.4-trillion-parameter model.[1][5]

Overview

FieldDetails
DeveloperAlibaba's Qwen team[1][5]
Production releaseNovember 16, 2026[2]
Production API identifierqwen4-max[3]
Earlier previewqwen4-max-preview, opened November 11, 2026[2][4]
Input / outputText, images and video / text[6]
ArchitectureSparse MoE; 2.4T backbone, 71B active per token; 328B additional N-gram Embedding parameters[5]
Advertised context length1,000,000 tokens[3]
QwenCloud input limits991,808 tokens without thinking; 983,616 with thinking enabled[3]
QwenCloud response limits131,072 tokens without thinking; 262,144 thinking and answer tokens combined with thinking enabled[3]
Reasoning settingslow, high, xhigh; default high; thinking can be disabled[7]
Standard Singapore API prices$1.50 input, $4.50 output and $0.075 cached input per million tokens[8]
Weight availabilityHosted proprietary model; smaller Qwen4 open-weight models were announced separately[1][9]

Expanded article table

Release and access

The November 11 preview and November 16 production model have different identifiers and access conditions. The preview was restricted to Token Plan, with no pay-as-you-go access. Its published preview period runs from 10:00 on November 11 through the end of November 25, 2026, in UTC+8, with retirement scheduled after that period. Alibaba recommends the production version.[4]

VersionAccess at introductionRelevant distinction
qwen4-max-previewToken Plan onlyTemporary preview; documentation lists a 393,216-token combined thinking-and-answer limit[4]
qwen4-maxPay-as-you-go API and Token PlanProduction model; QwenCloud lists a 262,144-token combined limit in thinking mode[3][6]

Expanded article table

QwenCloud documents OpenAI-compatible Chat Completions and Responses, Anthropic-compatible Messages, and DashScope access. The DashScope route uses its multimodal endpoint even for text-only requests. Token Plan uses its corresponding credentials and endpoint rather than treating subscription credits as ordinary pay-as-you-go billing.[3]

Regions and deployment scopes also matter. The service distinguishes its Singapore International offering from Chinese mainland and Global deployments. API keys are region-specific; a key must match the region of the endpoint. Global deployment can route inference across regions, so it is a deployment policy rather than a guarantee that processing stays in the endpoint's country.[10]

The announcement promised smaller open-weight models in the Qwen4 family. It did not announce that Qwen4-Max's weights were available for download.[1]

Architecture

The Qwen team's launch description identifies three changes to its sparse MoE design.[5]

ComponentPublished description
Gated DeltaNet and Qwen Sparse AttentionThree Gated DeltaNet layers retain compressed history in fixed-size states for each QSA layer; QSA selects context at micro-block granularity with a lightweight indexer.
Gated ResidualFour parallel residual branches use dynamic read and write gates to regulate information flow.
N-gram EmbeddingLocal token-pattern representations are retrieved at the second layer through context-dependent lookups.

Expanded article table

Alibaba says the active backbone count is 25% lower than Qwen3.8-Max at the same 2.4T backbone scale. These are vendor-reported architecture and training claims, rather than measurements from a released Qwen4-Max weight checkpoint.[1][5]

API capabilities

Visual input and tools

The QwenCloud model list documents image and video understanding alongside text generation. Its published visual-input limits include up to 16 million pixels per image, videos up to two hours or 2 GB, and up to 64 videos per request, subject to the model's token limits. It lists up to 2,048 images supplied by URL or 250 supplied as Base64 data. Audio in a video requires a model with audio support; QwenCloud directs that use case to its Omni models.[6]

Function calling, parallel calls, structured output, streaming and built-in tools are supported. The feature listing includes web search, PDF understanding, a code interpreter, web scraping, image search and Model Context Protocol tools. The code interpreter requires thinking mode. Batch inference, fine-tuning and managed model deployment are not supported in the documented QwenCloud offering.[3]

Thinking and output budgets

Thinking is enabled by default. QwenCloud's thinking guide specifies reasoning_effort=high, a default thinking_budget of 131,072 tokens, and preserved thinking history. Supported effort levels are low, high and xhigh; compatibility values such as medium map to high rather than introducing another effort level.[7]

In thinking mode, max_completion_tokens caps reasoning and the final answer together, up to 262,144 tokens. Without thinking, the documented maximum is 131,072. A budget sized only for the visible answer can therefore truncate a reasoning response.[3][7]

Caching

Qwen4's explicit context cache can reuse a specified input prefix through manual breakpoints or automatic breakpoint placement. It supports Chat Completions, Responses and Anthropic Messages, with five-minute and one-hour lifetimes. DashScope does not support this explicit-cache interface. A cached prefix must contain at least 1,024 tokens.[11]

Explicit cache hits require an identical prefix within its validity period and are reused across the same user's requests. Ordinary implicit caching automatically detects common prefixes but does not guarantee a hit. The API option called prompt_cache_options.mode="implicit" enables automatic breakpoint placement within the explicit-cache system; its name does not make it the same as ordinary implicit caching with no cache parameters.[11]

Reasoning summaries

The optional CoT Summary feature returns a readable reasoning summary together with encrypted full reasoning, replacing plaintext reasoning for that response. Clients must enable the feature on each turn and replay the encrypted reasoning unchanged when continuing a conversation. The summary is for display and cannot reconstruct the full reasoning history.[12]

Encryption covers thinking rather than the original images, videos or other inputs, so those inputs must be retained separately. Summaries can be incomplete, and ciphertext can be absent when no reasoning was generated or encryption failed. The documentation lists no extra feature charge beyond normal model token billing.[12]

Pricing

The following QwenCloud list prices were documented on November 17, 2026. Figures are US dollars per million tokens; output billing includes reasoning and answer tokens. Singapore International and Global deployments have different prices.[8]

Deployment scopeUncached inputOutputExplicit or implicit cache readExplicit cache write, 5 minutesExplicit cache write, 1 hour
Singapore, International$1.50$4.50$0.075$1.875$3.00
Global$1.413$4.24$0.071$1.767$2.827

Expanded article table

A cache-read price is not the price of initially creating an explicit cache. The listed rates also do not turn Token Plan credits into a universal per-token subscription rate. Prices, discounts and model availability must be checked for the selected region and service deployment scope.[8][10][11]

Evaluation results

Alibaba's Work360 evaluation

Work360 is the Qwen team's own professional-work evaluation, organized across 26 top-level domains, 182 subdomains and 721 specialized fields. The launch post reports a score of 52.69 for Qwen4-Max versus 47.74 for Qwen3.8-Max.[5]

Alibaba's press release reports 35% fewer tokens at High reasoning effort and an average task-cost reduction of around 60% on Work360. The broader social announcement uses an "up to" token-reduction figure across evaluations. Those claims describe the vendor's test configurations, rather than a guaranteed saving on a customer's workload.[1][13]

Arena's launch-day results

Arena announced the following preliminary WebDev Arena results on November 16, 2026.[14]

Arena categoryQwen4-Max ratingReported rank
Code Arena: WebDev1750, preliminary6
HTML17486
Brand & Marketing17058
Reference-Based Design17708

Expanded article table

The WebDev rating was 74 points above Qwen3.8-Max (0902), which Arena listed at 1676. These are dated ratings from a particular evaluation surface, not percentages of tasks solved or an overall intelligence ranking.[14]

Arena's separate Agent Arena announcement placed Qwen4-Max 14th overall on the same day. It reported a +7.5% net-improvement figure across 14.7K community agentic sessions, with Qwen3.8-Max ranked 39th. The announcement also reported these signal-specific results:[17]

Agent Arena signalReported net figureReported rank
Praise vs. Complaint+13%15
Confirmed Success+9%13
Steerability+8%19
Bash Reliability+3%29

Expanded article table

These are Arena's reported net figures, not absolute task-success rates. The announcement does not establish that every one of the 14.7K sessions evaluated Qwen4-Max alone.[17]

Artificial Analysis

Artificial Analysis listed Qwen4-Max at 43 on its Intelligence Index v5.0.1 when checked on November 17. Its page reported an estimated $3.27 per Intelligence Index task and 32.2 minutes per task. It also identified the model as proprietary and listed text, image and video input with text output.[9]

That index combines nine evaluations across knowledge work, coding, science and general knowledge, with each category contributing 25%. It is primarily an English-language, text-based suite, with images included in selected tasks. The methodology evaluates other visual, multilingual and speech capabilities separately. Its task costs and timing therefore describe its own evaluation mix rather than every possible use of the model.[15]

Long-running workflows

The release emphasizes AI agents that produce usable versions during an extended task. The accompanying Qwen-Live-Delivery-Plugin is a separate Apache-2.0 software project. Its launch README lists Codex CLI support on Linux and macOS with Node.js 22 or later; it does not bundle a model or API key.[16]

The plugin combines a delivery skill, an MCP service that records plans and version snapshots, and lifecycle hooks for handling feedback and task completion. A local browser workbench exposes previews, user feedback, the remaining time budget and a stop control. Its code license does not make Qwen4-Max's model weights open.[16]

Alibaba also demonstrated manufacturing, cloud scheduling and laboratory workflows. Its press release describes a 48-hour, four-iteration coolant-design experiment and a separate drug-design workflow producing 24 candidate molecules. These are demonstrations reported by the developer; computational drug candidates are not evidence of a clinically effective treatment.[1]

Safety and interpretation

Alibaba says it incorporated secure-coding checks into software-engineering training and filtered cybersecurity training examples by task intent, retaining defensive uses while removing offensive material. It also describes dedicated CBRN safety data and post-training safeguards. These statements report the developer's process; they do not establish immunity from unsafe outputs or vulnerabilities.[1]

The independent evaluations, vendor benchmarks and product demonstrations measure different things. Arena ratings reflect its evaluation surface, Work360 is proprietary, and Artificial Analysis applies its own benchmark mix and token accounting. Their figures should retain their dates and test settings when compared.[9][14][15]

Continue reading

More to read

Topics:Chinese AILarge Language ModelsMultimodal AIReasoning Models

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Alibaba Cloud. "Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery." November 16, 2026. alibabacloud.com/...alibaba-unveils-qwen4-max
  2. ^1 ^2 ^3QwenCloud. "Model releases," November 11 and November 16, 2026 entries. Accessed November 17, 2026. docs.qwencloud.com/...models
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9QwenCloud. "Latest model: Qwen4-Max." Accessed November 17, 2026. docs.qwencloud.com/...latest-model
  4. ^1 ^2 ^3 ^4Alibaba Cloud. "qwen4-max-preview." Accessed November 17, 2026. help.aliyun.com/...qwen4-max-preview
  5. ^1 ^2 ^3 ^4 ^5 ^6Qwen Team. "Qwen4-Max: Your New Teammate Has Arrived." November 16, 2026. qwen.ai/blog
  6. ^1 ^2 ^3QwenCloud. "Text generation and visual understanding." Accessed November 17, 2026. docs.qwencloud.com/...text-generation-models
  7. ^1 ^2 ^3QwenCloud. "Thinking," Qwen4-Max thinking configuration. Accessed November 17, 2026. docs.qwencloud.com/...thinking
  8. ^1 ^2 ^3QwenCloud. "Model pricing," Qwen-Max tables and cache prices. Accessed November 17, 2026. docs.qwencloud.com/...model-prices
  9. ^1 ^2 ^3Artificial Analysis. "Qwen4 Max: Intelligence, Performance & Price Analysis." Accessed November 17, 2026. artificialanalysis.ai/...qwen4-max
  10. ^1 ^2QwenCloud. "Regions, deployment scopes, and endpoints." Accessed November 17, 2026. docs.qwencloud.com/...regions
  11. ^1 ^2 ^3QwenCloud. "Explicit cache (Qwen4)." Accessed November 17, 2026. docs.qwencloud.com/...explicit-cache-qwen4
  12. ^1 ^2QwenCloud. "Thinking summary and encryption." Accessed November 17, 2026. docs.qwencloud.com/...cot-summary
  13. ^QwenCloud (@qwen_cloud). Qwen4-Max release announcement. X, November 16, 2026. x.com/...2122149748530675756
  14. ^1 ^2 ^3Arena.ai (@arena). Qwen4-Max Code Arena: WebDev announcement. X, November 16, 2026. x.com/...2122153582384583156
  15. ^1 ^2Artificial Analysis. "Intelligence Benchmarking Methodology," v5.0.1. Accessed November 17, 2026. artificialanalysis.ai/...intelligence-benchmarking
  16. ^1 ^2Qwen Team. "Qwen-Live-Delivery-Plugin." GitHub repository and README, accessed November 17, 2026. github.com/...Qwen-Live-Delivery-Plugin
  17. ^1 ^2Arena.ai (@arena). Qwen4-Max Agent Arena debut announcement. X, November 16, 2026. x.com/...2122153116959482051

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 2,086 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Cite this page: AI Wiki. "Qwen4-Max." aiwiki.ai, updated 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/qwen4_max

Edit article