Alibaba logo

Qwen4-Max

Multimodal
Alibaba

Qwen4-Max is Alibaba's next-generation flagship multimodal model in the Qwen family, generally available on 2026-11-16 (model id qwen4-max). It is a sparse mixture-of-experts model with 2.4 trillion backbone parameters, 71 billion active parameters per token, and an additional 328 billion N-gram Embedding parameters, with native text, image, and video understanding, a 1,000,000-token context window, and thinking enabled by default (reasoning_effort levels low, high, and xhigh; default high; thinking_budget default 131,072; maximum thinking plus final answer 262,144 tokens; maximum input 983,616 tokens in thinking mode and 991,808 in non-thinking mode). It supports function calling, web search, structured outputs, explicit and implicit cache, and the OpenAI Chat Completions and Responses APIs and the Anthropic Messages API. QwenCloud Singapore list pricing is $1.50 per million input tokens and $4.50 per million output tokens, with cache write $1.875 (5-minute TTL) or $3 (1-hour TTL) and cache read $0.075 per million tokens. Beijing and Global regions list $1.413 input and $4.24 output, with cache write $1.767 (5-minute TTL) or $2.827 (1-hour TTL) and cache read $0.071 per million tokens. This card is the generally available model; the earlier Token Plan-only preview is cataloged separately as qwen4-max-preview. The official Qwen team launch post says smaller models in the Qwen4 family are coming to open source; no open weights for Qwen4-Max were published at catalog time.

Key Specifications

Parameters
2.4T
Context
983.6K
Release Date
November 16, 2026
Average Score
61.9%

Timeline

Key dates in the model's history
Announcement
November 16, 2026
Last Update
November 18, 2026
Today
November 20, 2026

Technical Specifications

Parameters
2.4T
Training Tokens
-
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)
$1.50
Output (per 1M tokens)
$4.50
Max Input Tokens
983.6K
Max Output Tokens
262.1K
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks
Agents' Last Exam
Agents' Last Exam v1.1 score 46.7 for Qwen4-Max, read from the official Qwen team launch blog summary figure. Official footnote: scores are cited from the Agents' Last Exam. • Self-reported
46.7%
APEX-Agents
APEX-Agents v1.1 score 72.1 for Qwen4-Max, read from the official Qwen team launch blog summary figure (panel labeled Professional services tasks). Official footnote: scores are cited from the Mercor leaderboard. • Self-reported
72.1%
BioMysteryBench
BioMysteryBench score 85.0 for Qwen4-Max, read from the official Qwen team launch blog summary figure (panel labeled Bioinformatics research). The official figure and footnotes do not specify the Human Solvable or Human Difficult subset, so this uses the parent biomysterybench id. • Self-reported
85.0%
Chartography
Chartography score 54.2 for Qwen4-Max, read from the official Qwen team launch blog summary figure (panel labeled Visual chart understanding). Official footnote: scores for Qwen4-Max and Qwen3.8-Max-0902 were obtained with min_pixels set to 4096x32x32. The official figure and footnotes do not state tools vs no-tools, so this uses the parent chartography id. • Self-reported
54.2%
DeepSWE 1.1
DeepSWE v1.1 score 75.7 for Qwen4-Max, read from the official Qwen team launch blog summary figure. • Self-reported
75.7%
HealthBench Professional
HealthBench Professional score 60.0 for Qwen4-Max, read from the official Qwen team launch blog summary figure (panel labeled Clinical reasoning). • Self-reported
60.0%
Humanity's Last Exam
Humanity's Last Exam score 51.7 for Qwen4-Max, read from the official Qwen team launch blog summary figure (panel labeled Multidisciplinary reasoning). The official figure and footnotes do not state a tools or text-only setting, so this uses the generic humanity's-last-exam id. • Self-reported
51.7%
Job Bench
JobBench score 68.5 for Qwen4-Max, read from the official Qwen team launch blog summary figure (panel labeled Professional job tasks). Official footnote: scores were obtained using the updated dataset and evaluation code. Filed under job-bench, the id used for other Qwen models. • Self-reported
68.5%
Program Bench
ProgramBench score 31.5 for Qwen4-Max, read from the official Qwen team launch blog summary figure (panel labeled Black-box software replication). • Self-reported
31.5%
Terminal-Bench 4.0
Terminal-Bench 4.0 score 48.5 for Qwen4-Max, read from the official Qwen team launch blog summary figure (panel labeled Agentic terminal coding). • Self-reported
48.5%
WebDev Arena
Code Arena WebDev Elo of 1,750 for Qwen4-Max, as shown on the official Qwen team launch blog summary figure (https://qwen.ai/blog?id=qwen4-max), where the panel is labeled CodeArena (Elo) / Web development. The same 1,750 xHigh Elo, ranked 6th, appears on Alibaba Cloud's official launch Code Arena figure (source label: Arena AI leaderboard). • Self-reported
87.5%

License & Metadata

License
proprietary
Announcement Date
November 16, 2026
Last Updated
November 18, 2026

Compare Qwen4-Max

All comparisons

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.