Q
ALIBABA / QWEN / qwen4-max

Qwen4-Max

Alibaba multimodal model introduced on November 16, with a MoE architecture of 2.4 trillion total parameters and 71 billion active per token. It understands text, images and video and generates text for coding, visual analysis and agent tasks.

01 / SUMMARY

The essential

ORIENTATIONVisual reasoning, coding and agents

Primary use stated or inferred cautiously from official documentation.

CONTEXT1,000,000 tokens; maximum input of 991,808 without reasoning or 983,616 with it

Maximum input capacity when published by the source.

EXIT131,072 tokens without reasoning; 262,144 for reasoning and answer combined

Documented maximum generation limit.

ACCESSAlibaba Cloud Model Studio · API and Token Plan

Verified availability channels.

MODEL TYPEText and reasoning · Code · Multimodal

Primary functional family and verified specialties.

LICENSEProprietary

Declared terms for API access, model weights, or self-hosting.

02 / TECHNICAL SHEET

Limits and integration

API IDqwen4-max
Model typeText and reasoning · Code · Multimodal
Access modelPaid proprietary
LicenseProprietary
DeploymentHosted API
Release16 NOV 2026
Knowledge cutoffNot published
EntranceText · Image · Video
ExitText
Context window1,000,000 tokens; maximum input of 991,808 without reasoning or 983,616 with it
maximum output131,072 tokens without reasoning; 262,144 for reasoning and answer combined
ReasoningOptional; enabled by default at high effort. Supports low, high and xhigh.
Published toolsFunction calling · Web search · Parallel tools
Structured OutingsJSON Object and JSON Schema
Batch processingUnavailable
Prompt cacheImplicit and explicit caching; TTL of 5 minutes or 1 hour
Fine-tuningUnavailable
Verified platformsAlibaba Cloud Model Studio
BEST SUITED FOR

Work with long documents, develop interfaces and coordinate tools in multistep tasks. The API supports streaming, parallel calls and JSON formats through Chat Completions, Responses, Anthropic Messages and DashScope.

WORTH MONITORING

One million tokens is the context window, not the output limit. Reasoning uses part of the 262,144 generation tokens. Batch inference, fine-tuning and self-hosting are unavailable. Access, quotas and regional data handling depend on the account and contracted scope. Future smaller Qwen4 models are not part of this release.

03 / CAPABILITIES

What it can do

01

Orientation

Work with long documents, develop interfaces and coordinate tools in multistep tasks. The API supports streaming, parallel calls and JSON formats through Chat Completions, Responses, Anthropic Messages and DashScope.

02

Context

1,000,000 tokens; maximum input of 991,808 without reasoning or 983,616 with it · 131,072 tokens without reasoning; 262,144 for reasoning and answer combined

03

Tools and integration

Function calling · Web search · Parallel tools

04

Access

Alibaba Cloud Model Studio · API and Token Plan

Input modalities
TextImageVideo
04 / PRICES

Documented cost

ConceptWorthUnit/condition
Standard inputUSD 1.50 / 1M tokens · SingaporeStandard API rate
Cached inputNot publishedReading reused prefixes
Cache write or storageNot publishedThe condition varies by provider
Standard outputUSD 4.50 / 1M tokens · SingaporeMay include reasoning tokens
Batch inputNot publishedAsynchronous processing
Batch outputNot publishedAsynchronous processing

Singapore international pricing: USD 1.50 input and USD 4.50 output per million tokens, including reasoning tokens. Cache reads cost USD 0.075; creation costs USD 1.875 with a 5-minute TTL or USD 3 with a 1-hour TTL. Global scope lists USD 1.413 input and USD 4.24 output. Do not mix regions or apply batch discounts to this model.

i Prices change and may depend on level, region or context length. Check the source before making a decision.

05 / EVALUATIONS

How to read the results

“

The release publication shows 1,750 points and sixth place on Code Arena WebDev at xHigh effort, and 52.69 on Work360, a proprietary professional-task evaluation. These are published results, not tests reproduced by Inferama. The Arena ranking is a snapshot; Work360 and other tool-based tests depend on the environment and do not guarantee overall quality.

BenchmarkResultMetricSource
Code Arena WebDev1,750Arena score · xHighView source ↗
Work36052.69Overall score · proprietary evaluationView source ↗

Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.

06 / VERSIONS

Chronology

qwen4-max

Stable pay-as-you-go version; distinct from qwen4-max-preview, released on November 11.

07 / SOURCES

Traceability