← All AI resources

Frontier model release · Qwen4-Max launch and evaluation guide

Alibaba lowers Qwen4-Max prices, but buyers still need their own test

Alibaba released its closed Qwen4-Max model with a one-million-token context window, lower API prices and new claims for long-running agent work.[1][3] Plotly reports a more mixed result.[6]

Published by Learnetto team · Published November 16, 2026

Qwen4-Max adoption map

2.4T backboneAlibaba says 71B parameters activate per token
$1.50 / $4.50International input and output price per million tokens
Mixed evidencePlotly and Artificial Analysis report narrower results
Test one queueMeasure accepted work, corrections and whole-run cost

Qwen4-Max is closed today. Alibaba says smaller open-weight Qwen4 models will follow.

Alibaba released Qwen4-Max on November 16 as the first flagship model in the Qwen4 family. The company describes a mixture-of-experts model with 2.4 trillion backbone parameters and 71 billion active parameters per token.[1] Alibaba Cloud documents text, image and video inputs.[2] The international API charges $1.50 for a million input tokens and $4.50 for a million output tokens, 25% below Qwen3.8-Max at the same deployment scope. Cached input reads cost $0.075 per million tokens.[3] Alibaba reports broad gains in coding, professional work and long-running research. Plotly and Artificial Analysis do not support one universal ranking. Artificial Analysis places the model well behind the leading frontier systems on its composite index, while Plotly found that an older open-weight Qwen model scored higher and cost less on its analytics test.[6][7]

Released
November 16, 2026
Architecture
2.4T backbone · 71B active / token
Context
1,000,000 tokens
Weights
Qwen4-Max is closed

What Alibaba released

Qwen's launch post [1] introduces Qwen4-Max as a native multimodal model for professional work, software tasks and extended agent runs. Alibaba says the sparse model has 2.4 trillion backbone parameters and activates 71 billion for each token.

Alibaba Cloud documents text, image and video input with text output, tool calling, web search, structured output, prompt caching and thinking controls. The service accepts OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope requests.[2]

The context window is one million tokens. Alibaba lists a 131,072-token maximum output in non-thinking mode and a 262,144-token combined limit for reasoning plus the final answer in thinking mode.[2]

Alibaba lowered list prices, but task cost depends on behavior

Alibaba Cloud lists $1.50 for each million input tokens and $4.50 for each million output tokens in its international scope. Qwen3.8-Max costs $2 and $6 at the same scope, so each standard rate falls by 25%. Qwen4-Max cache reads cost $0.075 per million tokens, while five-minute and one-hour cache writes cost $1.875 and $3.[3]

Alibaba says Qwen4-Max used 35% fewer tokens than Qwen3.8-Max at high reasoning effort on its proprietary Work360 test and reduced average task cost by about 60%.[4] Those figures come from Alibaba's own benchmark, so a buyer should reproduce the comparison with the prompts, tools and acceptance checks that matter to the buyer.

You cannot infer completed-task cost from the token price alone. Long reasoning traces, repeated tool calls, retries and human corrections can cost more than the first response. Measure successful work and record the token price.

What early independent testers found

Artificial Analysis gives Qwen4-Max a score of 43 on its Intelligence Index and reports a weighted cost of $3.27 per benchmark task. Artificial Analysis lists 25 other model configurations above Qwen4-Max, including several Claude, Gemini and OpenAI variants.[7]

Artificial Analysis reports 72.6% on AutomationBench-AA, 48.5% on Terminal-Bench 4.0 and 26.5% on FrontierCode-AA 2.0.[7] When you rely on one composite score, you may miss whether a model fits coding, office automation, research or another specific workload.

Plotly ran one 43-question analytics exam and recorded 35 correct answers for $2.29. Its older open-weight Qwen 3.8 27B run answered 38 questions for $1.71. Plotly says repeat runs of other models have moved by as many as three points. Treat this result as one warning and wait for repeat runs before ranking the models.[6]

Alibaba keeps Qwen4-Max closed

Qwen4-Max is available through hosted products and APIs, but Alibaba has not released its weights. Plotly also notes that Alibaba has not published enough information to determine whether the company trained it from scratch or extended an earlier model.[6]

Alibaba says smaller Qwen4 models will come with open weights later.[4] Developers cannot deploy Qwen4-Max weights themselves today. Teams that require local inference, weight inspection or deployment outside Alibaba's service should compare models with weights available now.

Developers may spend less integration time with compatible APIs, but they must still handle provider-specific behavior. Alibaba's defaults include thinking mode, high reasoning effort and preserved reasoning. Test parameters, tool schemas, encrypted reasoning continuity and error handling before switching production traffic.[2]

What Alibaba says about safety

Alibaba says it included code-security checks during software-engineering training and targeted injection, memory-safety and broken-access-control defects. The company also says it filtered offensive cyber material and added training for chemical, biological, radiological and nuclear risks.[4] Reuters reported Alibaba's safety claims.[5]

Alibaba described its training choices, but buyers still lack a public guarantee for every generated program or agent action. Keep code review, sandboxing, least-privilege tools and deterministic policy checks around the model.

Alibaba does not provide enough detail in its public announcement to reproduce every safety result. Buyers should ask for applicable evaluation reports and run misuse, prompt-injection and tool-boundary tests in their own harness.

A practical Qwen4-Max evaluation

  • Select 20 recent tasks with known acceptance criteria, including several that require tools, images or long context if those capabilities matter.
  • Freeze the system prompt, tool definitions, repository or data snapshot, network permissions, time limit and retry policy for every model.
  • Run your current model and Qwen4-Max at documented reasoning settings at least twice when the budget permits.
  • Record passed checks, reviewer corrections, elapsed time, input tokens, reasoning and answer tokens, cache activity, tool calls, retries and total spend.
  • Test malformed tool output, prompt injection, unavailable services and context near the intended maximum before any production trial.
  • Route one bounded workload first, preserve a rollback model and define the quality or cost threshold that ends the trial.

What remains uncertain

Alibaba has not published the Qwen4-Max weights, complete training recipe or enough evidence for outsiders to reproduce Work360 and the physical-science demonstrations.[1][4][6]

The company reports a 48-hour coolant experiment that produced an experimentally validated coolant molecule and an Alzheimer's workflow that executed hundreds of tool calls.[4] Outsiders cannot yet verify independent replication, comparative baselines or the total laboratory and inference cost from the public material.

Early independent evaluators used narrow task sets, and evaluators may get different results when they change providers, prompts or harnesses. More repeated tests are needed before calling Qwen4-Max the best choice for a broad category.

Frequently asked questions

  • Is Qwen4-Max open weight? No. Alibaba says smaller Qwen4 models will receive open-weight releases later, but Qwen4-Max is currently a hosted proprietary model.[4]
  • How much does the API cost? In Alibaba Cloud's international scope, standard input costs $1.50 for each million tokens and output costs $4.50 for each million tokens. Regional prices differ.[3]
  • Can I call it with an OpenAI or Anthropic client? Alibaba documents compatibility with OpenAI Chat Completions, OpenAI Responses and Anthropic Messages, but you still need to test provider-specific parameters and responses.[2]
  • Does it beat Qwen3.8-Max? Alibaba reports stronger overall Work360 results with fewer tokens.[4] Plotly scored the older open-weight Qwen 3.8 27B higher at a lower cost in its separate analytics exam.[6]
  • Should I replace my coding model now? Run the same repository tasks with fixed tools and acceptance tests first. Artificial Analysis reports a much lower FrontierCode-AA result than Qwen4-Max's AutomationBench-AA result, so workload fit matters.[7]
  • What should I watch next? Watch for independent repeat evaluations, provider latency, the promised open-weight Qwen4 releases and any fuller safety or technical report.[4][6][7]

Keep learning on Learnetto

Sources

[1] Qwen: Qwen4-Max launch post

[2] Alibaba Cloud: qwen4-max model documentation

[3] Alibaba Cloud: Model Studio pricing

[4] Alibaba Cloud: Qwen4-Max press release

[5] Reuters via The Standard: Alibaba unveils Qwen4-Max

[6] Plotly: Qwen4-Max data analytics benchmark

[7] Artificial Analysis: Qwen4-Max evaluation results