Qwen4-Max: Specs, Pricing, Benchmarks & API

By Versus Tools AI Team · Published · Updated · 9 min read · AI Models & News

We may earn a commission if you buy through links on this page. Read the affiliate disclosure · How we review tools

Quick answer. Qwen4-Max launched Nov 16, 2026. Specs, verified API pricing, benchmarks versus Qwen3.8-Max, and how to call it.

Qwen4-Max is Alibaba’s new flagship, released Nov 16, 2026: a 2.4 trillion-parameter mixture-of-experts model that activates 71 billion parameters per token, with native vision and a 1 million-token context. The API id is qwen4-max. In Singapore the list price is $1.50 per million input tokens and $4.50 per million output tokens. The $0.075 rate in the launch post is the implicit cache-hit input price, not the price of a normal prompt.

This page is based on official documentation, benchmarks and public reports. We did not call the API or run Work360.

How this page was checked

We read Alibaba’s Nov 16, 2026 announcement, the Model Studio release log, the qwen4-max model doc (last updated Nov 16, 2026), Model Studio’s inference pricing page, and QwenCloud’s pricing page. Rival prices come from each vendor’s own pricing page, checked the same day. If a cell is not on those pages, the table says “not published”.

Google’s fast model from four days earlier is Gemini 4 Hydrogen. The frontier model Google made generally available on Nov 10, 2026 is Gemini 4 Carbon. Anthropic’s Nov 9 release is Claude Fable 5.2. These launches are on the AI model release timeline and in the best AI models ranking.

Key facts

Qwen4-Max
ReleasedNov 16, 2026. Preview id qwen4-max-preview is on the Model Studio log for Nov 11, 2026, Token Plan only.
MakerAlibaba. The announcement dateline is Hangzhou.
API model idqwen4-max
ArchitectureSparse mixture-of-experts. 2.4 trillion total parameters. 71 billion active per token, which Alibaba says is 25 percent fewer active parameters than Qwen3.8-Max at the same 2.4 trillion scale.
ModalitiesText, image, and video in. Text out. The announcement calls this native vision.
Context1,000,000 tokens.
Max output131,072 tokens in non-thinking mode. 262,144 tokens combined for thinking and the answer. Max input is 991,808 tokens non-thinking and 983,616 tokens in thinking mode.
Default effortreasoning_effort defaults to high. Supported values on the model doc are low, high, and xhigh.
Standard price, Singapore$1.50 / 1M input tokens, $4.50 / 1M output tokens. No input-length tiers.
Cache read, Singapore$0.075 / 1M tokens, explicit or implicit. Cache writes: $1.875 for 5 minutes, $3 for 1 hour.
Beijing / Global price$1.413 input and $4.24 output per million tokens on QwenCloud. Cache read $0.071.
Where it runsQwenCloud via the Token Plan and the API. Also QwenWork and QwenStudio.
Open weightsNot for Qwen4-Max. Smaller Qwen4 open-weight models are described as coming soon.

What Alibaba says changed

The announcement calls Qwen4-Max the most capable Qwen model so far, aimed at work, creativity, and research. The architecture note is specific: the sparse mixture-of-experts design was changed across attention, residual connections, and embeddings, and the active-parameter count dropped 25 percent versus Qwen3.8-Max while the total stayed 2.4 trillion.

The benchmark Alibaba uses for that claim is Work360, which it describes as a proprietary test of 721 specialized fields. On it, Qwen4-Max beats Qwen3.8-Max overall, uses 35 percent fewer tokens at High effort, and lowers average cost per task by about 60 percent. Work360 is Alibaba’s benchmark. We do not have a public paper for it in the launch post, so treat the margin as a company result.

Alibaba also says Qwen4-Max ranked 6th on Code Arena. The post does not print the score, the date of the board, or the five models above it. Coding, agent tasks, and visual reasoning are the areas it highlights, with Agents’ Last Exam, Vals Index, JobBench, and Chartography named and no scores attached. Those names without numbers stay out of the comparison table.

Two product examples in the post are also company-reported. A cloud scheduling test cut data-annotation model-call cost by 9 percent on average and sped offline CPU rescheduling by more than 14 times. A machining showcase cut workpiece programming from about 475 minutes to 24 minutes. A research example describes a 48-hour agent run that proposed a coolant molecule, and an Alzheimer’s workflow with 454 tool calls. Those are illustrations from Alibaba, not independent replications.

Pricing and the $0.075 line

The launch post says: input on an implicit cache hit is $0.075 per million tokens. Model Studio’s Singapore price card shows why that number is easy to misread. The standard input price is $1.50. The $0.075 rate is the cache read. You pay $1.50 when the prompt misses the cache.

Singapore (international), from Model Studio’s pricing page, for thinking and non-thinking, any request up to 1 million input tokens:

Billing itemPer 1M tokens
Input$1.50
Output, chain of thought plus the answer$4.50
Cache read, explicit or implicit$0.075
Cache write, 5 minutes$1.875
Cache write, 1 hour$3
Free quota1 million tokens, Singapore only, valid 90 days from activation, release, or approval, whichever is later

QwenCloud’s pricing page prints a second column for Beijing and for Global deployment (Frankfurt, Hong Kong, Tokyo, and US Virginia): input $1.413, output $4.24, 5-minute cache write $1.767, 1-hour cache write $2.827, cache read $0.071. Model Studio’s China (Beijing) table matches the $1.413 / $4.24 pair. Batch inference is listed as not supported for qwen4-max. There is no context-length tier.

Qwen3.8-Max, on the same Singapore card, is $2 per million input tokens and $6 per million output tokens, also flat to 1 million tokens, with a 1 million-token free quota. The lower Qwen4-Max list price is part of how Alibaba gets to a cheaper task. The 35 percent token cut is the other part. We are not combining them into a new estimate. The “about 60 percent” cost cut is Alibaba’s Work360 figure.

The Nov 11 preview, qwen4-max-preview, was Token Plan only. The Nov 16 model is the API id qwen4-max. The announcement says Qwen4-Max is on the Token Plan and the API. The Token Plan’s own allowance is not on the pages we priced, so we are not quoting one.

How to access it

Alibaba documents three product paths and six deployment regions. It does not, on these pages, publish a separate “outside China” restriction beyond which region you call.

  • Web and apps: QwenWork and QwenStudio, named in the announcement.
  • API: model id qwen4-max on Model Studio / QwenCloud. Protocols listed on the model doc: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and DashScope.
  • Regions on the model doc: Beijing for the Chinese mainland; Singapore for international; Frankfurt, Virginia, Tokyo, and Hong Kong for Global. Capabilities listed in each region match: text, image, and video input, text output, streaming, thinking, parallel tool calls, web search, structured output, and explicit plus implicit cache. Fine-tuning and self-deployment are marked unsupported.
  • Open weights: the announcement says smaller open-weight Qwen4 models will follow for developers worldwide. Qwen4-Max is not that release.

Default sampling on the model doc is temperature 1.0 and top_p 0.95 in thinking mode, and temperature 0.7 and top_p 0.8 in non-thinking mode. Thinking is on by default. If you send medium or max on an OpenAI-style effort parameter, the doc maps medium to high and max to xhigh.

👍 Pros

  • 71B active parameters, 25 percent fewer than Qwen3.8-Max, at a lower Singapore list price of $1.50 / $4.50
  • Documented in Beijing, Singapore, and four Global regions, with OpenAI-compatible and Anthropic-compatible APIs
  • Cache-hit input at $0.075 per million tokens in Singapore

👎 Cons

  • The headline $0.075 rate is a cache hit. A cache miss is $1.50 in Singapore
  • Code Arena rank and Work360 gains are Alibaba's, without a public score card in the post
  • Qwen4-Max weights are not open. Smaller open-weight Qwen4 models are still described as upcoming

Qwen4-Max vs Qwen3.8-Max and rivals

Prices are list rates on Nov 16, 2026. DeepSeek’s peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Other hours, including weekends, are off-peak. “Not published” means that vendor’s pages, as we read them, did not print the cell.

Qwen4-MaxQwen3.8-MaxKimi K3.1 ProDeepSeek V4.1 FlashGemini 4 Hydrogen
ReleasedNov 16, 2026not published as a single date. Snapshot id qwen3.8-max-2026-09-02 is on Model StudioOct 24, 2026not published on DeepSeek’s pricing page. The served version is DeepSeek-V4.1-Flash-1017Nov 12, 2026
API idqwen4-maxqwen3.8-maxkimi-k3.1-prodeepseek-flashgemini-4-hydrogen
Parameters2.4T total, 71B active2.4T total. Active count not published. Alibaba says Qwen4-Max uses 25 percent fewer active parameters2.8T total, 104B active on the Hugging Face model card. The launch post says 3T-classnot publishednot published
Context1,000,0001M on the 0902 snapshot note1,048,5761M1M
VisionText, image, and video inNative vision, per the 0902 snapshot noteText and image on the model cardSupportedNative multimodal input, per Google
Input / 1M$1.50 Singapore. $1.413 Beijing / Global$2 Singapore. $1.65 Beijing$2.00Cache miss $0.15 off-peak, $0.30 peak$1.50
Output / 1M$4.50 Singapore. $4.24 Beijing / Global$6 Singapore. $4.951 Beijing$10.00$0.60 off-peak, $1.20 peak$7.50
Cache-hit input / 1M$0.075 Singapore. $0.071 Beijing / Globalnot published$0.20$0.003 off-peak, $0.006 peak$0.075, implicit only
Code Arena6th (Alibaba). Score not printednot publishednot publishednot publishednot published
Work360Better overall than Qwen3.8-Max. 35 percent fewer tokens at High. About 60 percent lower average cost per task (Alibaba)Baseline in that testnot publishednot publishednot published
Terminal-Benchnot publishednot published65.7 percent on Terminal-Bench 4.0 (Moonshot)not published57 percent on Terminal-Bench 4.1 (Google)

Terminal-Bench 4.0 and 4.1 are different tests. Kimi’s 65.7 percent and Hydrogen’s 57 percent should not be read as a head-to-head. Moonshot’s Oct 24, 2026 post also quotes GPT-6 Astra at 58.2 percent and Claude Opus 5.5 at 66.4 percent on its Terminal-Bench 4.0 run. Those are Moonshot’s citations.

Kimi’s standard rates are from the Kimi API pricing page. Highspeed, kimi-k3.1-pro-highspeed, is $4 input and $20 output per million tokens. DeepSeek’s rates are from DeepSeek’s API pricing page. Gemini’s rates are from the Gemini Developer API pricing page and are written up in Gemini 4 Hydrogen.

Which one to call

Call qwen4-max if you want Alibaba’s current flagship, the 1 million-token window, image and video input, and the Singapore price of $1.50 / $4.50, and you can use cache hits at $0.075. Stay on qwen3.8-max only if you have a measured workflow you do not want to move. On list price, the newer model is cheaper in Singapore.

Call deepseek-flash if the bill dominates and DeepSeek’s off-peak rates are acceptable. Call kimi-k3.1-pro if you want Moonshot’s open-weight flagship and will pay $2 / $10. Call gemini-4-hydrogen if you are already on Google and want the model Google priced at $1.50 / $7.50 for coding agents. None of those sentences is a quality ranking. The public numbers do not line up on one benchmark.

Qwen4-Max is a model. These are separate products for the work around it.

  • Writesonic if you want a writing app rather than a raw API.
  • Fireflies for meeting notes, which the model will not record for you.
  • ElevenLabs if the creative job is voice, not text or images.
See Writesonic →

Sources

Checked Nov 16, 2026.

Pricing last checked Nov 16, 2026. This page contains affiliate links. A commission does not change which API is described.

Frequently asked questions

When was Qwen4-Max released?

Alibaba's announcement is dated November 16, 2026. Alibaba Cloud Model Studio's release log lists the model id qwen4-max on the same day for international deployment. A preview id, qwen4-max-preview, is logged on Nov 11, 2026, and that preview was Token Plan only.

How much does the Qwen4-Max API cost?

On Alibaba's Model Studio pricing page for Singapore, qwen4-max is $1.50 per million input tokens and $4.50 per million output tokens, with no input-length tiers. Explicit and implicit cache reads are $0.075 per million tokens. The launch post's $0.075 figure is that cache-hit input rate, not the standard input rate. QwenCloud lists Beijing and Global deployment at $1.413 input and $4.24 output per million tokens, with cache reads at $0.071.

What is the Qwen4-Max API model name?

The Model Studio model doc says the invocation id is qwen4-max. The same doc says the API is compatible with OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and DashScope.

How does Qwen4-Max compare with Qwen3.8-Max?

Alibaba says both are 2.4 trillion-parameter mixture-of-experts models, and that Qwen4-Max activates 71 billion parameters per token, 25 percent fewer than Qwen3.8-Max. On Work360, Alibaba's own benchmark of 721 fields, Qwen4-Max is better overall, uses 35 percent fewer tokens at High effort, and costs about 60 percent less per task on average. Singapore list price is $1.50 / $4.50 for Qwen4-Max and $2 / $6 for Qwen3.8-Max.

Can I use Qwen4-Max outside mainland China?

Yes, on the regions Alibaba documents. The model doc lists Chinese mainland deployment in Beijing, international deployment in Singapore, and Global deployment in Frankfurt, Virginia, Tokyo, and Hong Kong. Each scope has its own published token price. This article does not add access rules beyond those pages.

Is Qwen4-Max open weight?

Not this model. Alibaba says smaller open-weight Qwen4 models will be released soon. Qwen4-Max itself is the hosted flagship on QwenCloud, QwenWork, and QwenStudio.