Robots Atlas>ROBOTS ATLAS
Qwen3-Max
AI Modelsโ€บQwen3

Qwen3-Max

Maxย ยทย Family: Qwen3
Flagship proprietary Qwen3 model - a MoE with over 1 trillion parameters trained on ~36T tokens. Available via Alibaba Cloud API and qwen.ai, a frontier-class competitor to GPT-5, Claude, and Gemini.
โœ“ Activeโœ“ Public accessโ˜… FeaturedLLMReasoning modelTool-using model๐Ÿ“ Qwen3
Context window
262K
tokens
Parameters
1T+ (MoE)
parameters
Max output
32,768
tokens
Release date
24 September 2025
Access:APIHostedDeployment:โ˜ Cloud

Overview

Qwen3-Max is Alibaba Cloud's flagship proprietary large language model, officially announced on September 24, 2025 (preview in July 2025 at the World AI Conference in Shanghai). It is the first Qwen3 model to exceed 1 trillion parameters (~1T). Its architecture is Mixture-of-Experts - activating only a subset of experts per query, which reduces real inference cost while retaining model scale. The model was pretrained on ~36 trillion tokens across 119 languages, with an extensive post-training stage.

Key capabilities: advanced reasoning (Qwen3-Max-Thinking), agentic abilities (planning, tool use, long-horizon tasks), coding (SWE-Bench Verified: 69.6% for the Instruct variant, competing with GPT-5 and Claude Opus 4), mathematics (strong AIME 2025 scores), long-context understanding up to 262,144 tokens (256K), multilingual instruction following. Alibaba positions the model as 'thinking-native' - able to decide automatically when to invoke extended reasoning and when to answer directly.

Qwen3-Max family variants: Qwen3-Max-Instruct (base, instruction-following), Qwen3-Max-Thinking (reasoning with extended chain-of-thought, can produce multimedia outputs - image, video - through integration with Qwen3-VL/Qwen3-Omni). Alibaba serves the model exclusively via API (DashScope / Alibaba Cloud Model Studio) and via the qwen.ai chatbot. Weights are not published - Qwen3-Max is one of the few closed Qwen models, contrasting with the broader Qwen3 family released mostly under Apache 2.0.

Results and positioning: Alibaba announced that Qwen3-Max achieves results comparable to or better than Claude Opus 4, Kimi K2, DeepSeek V3.1, and GPT-5 on selected agentic benchmarks (Tau-Bench, BFCL v3). Alibaba Cloud pricing: approximately USD 1.20 per 1M input tokens and USD 6.00 per 1M output (tiered by context length). The Qwen mobile app had 234 million users in May 2026 (Reuters).

Wider context: Qwen3-Max is central to Alibaba's proprietary frontier model strategy. Alibaba Cloud director Eddie Wu (Alibaba CEO, also leading the 'Alibaba Token Hub' AI Business Unit since March 2026) and Chief AI Architect Zhou Jingren (formerly head of Tongyi Lab) oversee development. Former Qwen head Junyang Lin left in March 2026 after the Qwen3.5 release. Wikidata: Q130234299. Official model page: qwen.ai/blog?id=qwen3-max.

Classification
LLMReasoning modelTool-using model
Family: Qwen3
Access & deployment
APIHosted
Cloud
Weights: Closed
Key parameters
๐Ÿ“ Context: 262K
๐Ÿงฉ Parameters: 1T+ (MoE)
โœ“ Tools
๐Ÿ“ฅ Input: text, image, documents, structured data

Technical specification

Context window
262K
tokens
Parameters
1T+ (MoE)
parameters
Max output tokens
32,768
tokens per response
Knowledge cutoff
1 Jun 2025
Knowledge boundary
License
Proprietary (Alibaba Cloud commercial)
Hardware requirements
The model is available exclusively via the Alibaba Cloud API (DashScope, Model Studio) - no weights available for self-hosting. Alibaba serves the model on its own GPU/ASIC infrastructure (including its own Hanguang / Pingtouge accelerators).
Features:โœ“ Tool use
Modalities
โฌ‡ Input
textimagedocumentsstructured_data
โฌ† Output
textcodestructured_data

Capabilities and applications

Native model capabilities
Advanced reasoning
The ability to perform multi-step, structured reasoning: analysing problems, planning steps, and drawing conclusions from hypotheses. Reasoning-first models (e.g. GPT-5.1 Thinking) dedicate a portion of inference to chains of thought before responding.
Category: reasoning
Extended thinking mode
A reasoning-model variant with a larger inference budget: more thinking cycles, higher answer precision at the cost of response time. Choice between 'standard' and 'extended' thinking is left to the user (e.g. the selector in GPT-5.2 Pro).
Category: reasoning
Adaptive reasoning effort
The model decides how much 'thinking' to allocate to a given query: simple questions are answered quickly, complex problems receive more inference cycles. A GPT-5.1 feature (both Instant and Thinking) that shortens time on easy tasks and extends it for hard ones.
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Mathematical reasoning
The model's ability to solve mathematical tasks requiring multi-step reasoning โ€” equations, proofs, combinatorics, geometry, calculus and competition-level problems.
Category: reasoning
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Agentic coding
Multi-hour, multi-step programming tasks performed autonomously by the model: cloning a repository, running tests, iterating on fixes, integrating with CLI tools. Characteristic of Codex variants (GPT-5.1-Codex-Mini, Codex-Max).
Category: coding
Agentic capability
The model's ability to autonomously plan and execute multi-step tasks by sequentially using tools, maintaining context, and adapting to intermediate results.
Category: planning
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language

Benchmark results

4 benchmarks
SWE-Bench Verified
resolved rate ยท Qwen3-Max-Instruct, agentic real-world GitHub issue fixing
69.6%%
๐Ÿ“… 24 Sept 2025๐Ÿ“„ Qwen3-Max Blog (qwen.ai/blog?id=qwen3-max, 2025-09-24)
Tau-Bench
pass rate ยท Agentic evaluation in airline / retail environments
N/A (comparable to Claude Opus 4)%
๐Ÿ“… 24 Sept 2025๐Ÿ“„ Qwen3-Max Blog (qwen.ai/blog?id=qwen3-max, 2025-09-24)
BFCL v3
function calling accuracy ยท Berkeley Function Calling Leaderboard v3, tool-use and function calling ability
N/A (top-tier)%
๐Ÿ“… 24 Sept 2025๐Ÿ“„ Qwen3-Max Blog (qwen.ai/blog?id=qwen3-max, 2025-09-24)
AIME 2025
accuracy ยท American Invitational Mathematics Examination 2025 competition, Qwen3-Max-Thinking variant
N/A (top-tier reasoning)%
๐Ÿ“… 1 Nov 2025๐Ÿ“„ Qwen3-Max-Thinking Blog (qwen.ai)

Pricing

Technical architecture

Core Architecture
TRTransformerMOMoE

Deployment and security

๐Ÿ”’ Security / Enterprise
โœ“ Verified enterprise information

The model is hosted exclusively in Alibaba Cloud - enterprise customers can select regions (China, Singapore, Malaysia, Germany, USA, Dubai) to meet data residency requirements. Alibaba Cloud holds ISO 27001, SOC 2, PCI DSS certifications. Due to its China base, the model is subject to Chinese AI regulations (including registration with the Cyberspace Administration of China). Weights are not available for self-hosting, which prevents offline deployment.

โ†— Security documentation

Sources and related pages