HomeAI models
Qwen: Qwen3 Max Thinking
Qwen
•
Qwen3 Max Thinking
•
Released 
January 2026

Qwen
Qwen3 Max Thinking

Alibaba's flagship reasoning model: 262K context, up to 81,920 thinking tokens, and strong tool-assisted HLE results.

Modality:
Text
model ID
qwen/qwen3-max-thinking
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

21.3
/ 100

Context Window *

262144
tokens

Input price

4.68
Anytoken

Output price

23.4
Anytoken
Qwen3 Max Thinking: Alibaba's Deliberate Reasoning Flagship for Tool-Assisted Problems Qwen3 Max Thinking is Alibaba's flagship reasoning model from the Qwen team, announced on 25 January 2026 as the thinking counterpart to Qwen3-Max. It is a closed-weight, text-in, text-out model with a 262,144-token context window and a chain-of-thought budget of up to 81,920 tokens. Alibaba's own results show its strongest gains on tool-assisted research questions, where it reports 49.8 on HLE with tools. It suits multi-step analysis, math and code reasoning, and agent loops where answer quality matters more than speed. Access Qwen3 Max Thinking through the AnyAPI.ai API

Performance

Where Qwen3 Max Thinking Gains Most: Reasoning That Can Call Tools

Qwen3 Max Thinking is strongest where reasoning can lean on tools. Alibaba reports 49.8 on Humanity's Last Exam with tools, ahead of the 45.5 it lists for GPT-5.2-Thinking. Without tools it reports 30.2, below GPT-5.2-Thinking at 35.5 and Gemini 3 Pro at 37.5. That gap suggests retrieval and code execution do much of the work on the hardest questions. In production, pair it with search or code tools for research and analysis, and budget for long reasoning traces rather than expecting fast responses.

Benchmarks

Independent Results: Real Reasoning Gains, Weak Knowledge Reliability

Artificial Analysis scored Qwen3-Max-Thinking 40 on its Intelligence Index at launch, up from 32 for the preview, but still behind peers. Its HLE result roughly doubled to 26%, ahead of DeepSeek V3.2 (22%) and GLM-4.7 (25%). The weak spot is the AA-Omniscience Index at -34, which measures accuracy and hallucination together and trails Kimi K2.5 (-11) and DeepSeek V3.2 (-23). It used about 86M output tokens, mostly reasoning, to finish the index. Compare scores only within one index version.

Output Speed

*
N/A
tok/s

Intelligence Index

*
21.3
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
86
%

HLE *

Adherence to multi-step structured instructions.
28
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Qwen3 Max Thinking takes text in and returns text, with a 262,144-token context window and up to 258,048 input tokens. The limit with the most production impact is output. In thinking mode, Alibaba documents a 32,768-token final answer cap, separate from a chain-of-thought allowance of up to 81,920 tokens. Reasoning tokens therefore consume budget and latency before the answer begins. Plan prompts and thinking budgets together, and do not assume the full context window is available for long generated documents.
Verified Specifications — 
Qwen3 Max Thinking
*
Input modalities
Text
output modalities
Text
Context window
262144
 tokens
Maximum output tokens
32768
Reasoning
Yes
Knowledge cutoff
January 2026
Pricing (standard)
4.68
 AnyTokens in
 / 
23.4
 AnyTokens out

Limitations & Trade-offs

Where Qwen3 Max Thinking falls short

1
Weak factual reliability. Artificial Analysis scored it -34 on the AA-Omniscience Index, which penalizes hallucination alongside accuracy, trailing Kimi K2.5, DeepSeek V3.2 and MiniMax-M2.1. Closed-book question answering, compliance lookups and anything that cites facts from memory are affected. Ground it with retrieval or tools, validate claims against sources, or prefer a model with a stronger knowledge score when unsupported facts carry real risk.
2
Output cap and reasoning overhead. In thinking mode, Alibaba documents a 32,768-token final answer limit, with up to 81,920 tokens of reasoning before it. Artificial Analysis recorded about 86M output tokens, 79M of them reasoning, to complete its index. Long report generation, high-volume classification and latency-sensitive chat are poor fits. Independent speed data for the final release was not available when this page was prepared, so measure latency on your own prompts.
3
Text-only and legacy status. The model accepts text and returns text, so screenshots, scanned documents and video need another model or a separate extraction step. Alibaba also lists the Qwen3 family as legacy, recommends Qwen3.8-Max, and documents that successor with a 1M-token window against 262K here. Workloads needing very long context, multimodal input or the vendor's current roadmap should evaluate the successor first.
4
Not the leader on agentic coding. In Alibaba's own comparison, Qwen3-Max-Thinking scores 75.3 on SWE-bench Verified, below GPT-5.2-Thinking at 80.0 and Claude Opus 4.5 at 80.9, and 85.9 on LiveCodeBench v6 against Gemini 3 Pro at 90.7. These are vendor-run results. For autonomous repository-level bug fixing, test a coding-focused alternative before committing.

Best-Fit Workloads

Where this model earns its place

01

Tool-assisted research analysis


Alibaba's reported 49.8 on HLE with tools, versus 30.2 without, shows the model gains most when it can retrieve information and run code. Adaptive tool use, covering search, memory and a code interpreter, was a headline feature of the launch. It suits research assistants and analyst workflows that gather evidence before answering. Because independent knowledge scores are weak, require source citations and verify retrieved claims.

02

Hard math and STEM problem solving


Alibaba reports 98.0 on HMMT Feb 25, 87.4 on GPQA and 85.9 on LiveCodeBench v6, all vendor-run. The 81,920-token thinking budget gives room for long derivations. This fits solution generation for tutoring products, scientific analysis, and offline verification of difficult problems. It is a poor fit for interactive use where seconds matter, since reasoning tokens arrive before the answer.

03

Long-document reasoning


A 262,144-token window with up to 258,048 input tokens lets the model reason across large contracts, specifications or multi-file code excerpts in one request. This suits cross-document reconciliation and structured findings reports. The thinking-mode answer cap of 32,768 tokens limits very long outputs. Inputs beyond roughly 258K tokens need chunking or a larger-context model such as Qwen3.8-Max.

04

Escalation tier for agent pipelines


Function calling plus a configurable thinking budget make it usable as the planner for hard cases in an agent system. Route routine steps to a lighter model and send ambiguous or multi-step tasks here. Its long reasoning traces and modest independent speed expectations make it a poor default for every step, so reserve it for decisions where reasoning quality outweighs latency.

Pricing in anytokens via AnyAPI
Input
4.68
₳
Output
23.4
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Qwen3 Max Thinking via AnyAPI.ai

Access Qwen3 Max Thinking through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Qwen3 Max Thinking without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Qwen3 Max Thinking and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Qwen3 Max Thinking against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Qwen3 Max Thinking from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Qwen3 Max Thinking for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Qwen3 Max Thinking has a 262,144-token context window, with up to 258,048 input tokens per Alibaba Model Studio documentation. In thinking mode the final answer is capped at 32,768 tokens, and the chain-of-thought can use up to 81,920 tokens. Context window and output limits are separate, so long prompts do not guarantee long answers.

Alibaba's Qwen team announced Qwen3 Max Thinking on 25 January 2026; the API snapshot is qwen3-max-2026-01-23. It remains available, but Alibaba now lists the Qwen3 family as legacy and recommends Qwen3.8-Max for new projects. It is a closed-weight model, so it is accessed through APIs rather than self-hosted.

Alibaba reports 87.4 on GPQA, 85.9 on LiveCodeBench v6, 75.3 on SWE-bench Verified and 49.8 on HLE with tools. Artificial Analysis independently scored it 40 on its Intelligence Index at launch, behind peers, with a weak -34 on AA-Omniscience. It is strong on tool-assisted reasoning but not the leader in coding or factual reliability.

No. Qwen3 Max Thinking accepts text input and produces text output, according to Artificial Analysis and Alibaba's Model Studio configuration. For screenshots, scanned pages or video, use a vision-capable model or extract the text first. This is a model-specific limit, not one shared by every model in the wider Qwen lineup.

Qwen3 Max Thinking is the reasoning variant, while Qwen3 Max is the standard model without extended thinking by default. Artificial Analysis rates the thinking version as more intelligent. The trade-off is reasoning overhead: it used about 86M output tokens, mostly reasoning, to complete the Intelligence Index. Choose Qwen3 Max for fast, simple tasks and the thinking version for hard problems.

* Benchmark data source: Artificial Analysis artificialanalysis.ai