Qwen3-Max

// Overview

Qwen3-Max is the largest model in Alibaba's Qwen family, crossing the 1-trillion-parameter mark with a Mixture-of-Experts design that activates only a fraction of those parameters per token. It launched in September 2025 — first as a preview on September 5, then as a full production release later that month — and is positioned as the highest-performing tier of the Qwen-Max line for complex, multi-step tasks. Unlike most of Alibaba's Qwen releases, Qwen3-Max is closed-weight: it runs only through Alibaba Cloud's Model Studio API, Qwen Chat, and third-party gateways such as OpenRouter, with no local deployment.

The model was pretrained on roughly 36 trillion tokens (about twice the data used for Qwen2.5) with a multilingual, coding, and STEM/reasoning emphasis. It ships in two flavors: Qwen3-Max-Instruct, a fast non-thinking model aimed at coding and agentic tool use, and Qwen3-Max-Thinking, a reasoning variant that can call a code interpreter and use parallel test-time compute. Through the API it exposes a 262,144-token context window and up to 32,768 output tokens.

Qwen3-Max is a text-only model — vision tasks in the Qwen3 generation are handled by separate Qwen3-VL models. On release, the preview version of Qwen3-Max reached the global top three on the LMArena text leaderboard, and Alibaba reported strong day-one results on coding and agentic benchmarks. Treat the most aggressive math claims as vendor-reported rather than independently verified.

// Benchmarks

1. SWE-bench Verified (Qwen3-Max-Instruct)69.6%
2. Tau2-bench (Qwen3-Max-Instruct, agentic tool-calling)74.8%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Tiered by input length on Alibaba Cloud Model Studio (International deployment). Batch invocation is charged at 50% of real-time pricing. Global-deployment pricing is lower.

Pricing source ↗

// Strengths

• Trillion-parameter MoE scale with efficient sparse activation per token
• Strong software-engineering performance (69.6 on SWE-bench Verified, Instruct variant)
• Top-tier agentic tool-calling (74.8 on Tau2-bench, Instruct variant)
• Large 262K-token context window for long documents and codebases
• Two variants — fast Instruct and reasoning-focused Thinking — under one model line
• Native function calling, structured JSON output, and web-search support via Model Studio

// Best for

• Agentic workflows and multi-tool orchestration
• Software engineering and code generation across large repositories
• Long-context document analysis and summarization
• Complex multi-step reasoning and STEM problem solving
• Multilingual text generation and assistance
• Building API-driven assistants with structured output and function calling

// How to access

// Qwen-Max — every version

The full lineage of the Qwen-Max line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.