Qwen3 Max

Qwen3 Max – Alibaba's Largest MoE Foundation Model

Qwen3 Max is Alibaba's largest and most capable language model to date, developed by the Qwen team under Alibaba Cloud. It is a Mixture-of-Experts (MoE) model with over 1 trillion parameters, pretrained on approximately 36 trillion tokens — roughly double the pretraining data used for its predecessor, Qwen2.5.

Two Deployment Variants

Qwen3 Max ships in two distinct variants with different operational profiles:

• Qwen3-Max-Instruct — a production-ready non-thinking model optimized for coding, instruction following, agent tool-calling, and multilingual tasks. It is OpenAI API-compatible and available via Alibaba Cloud Model Studio (qwen3-max) and Qwen Chat.
• Qwen3-Max-Thinking — a reasoning-focused variant that uses a code interpreter and parallel test-time compute scaling. When augmented with tool use, it is documented as achieving perfect scores on AIME 25 and HMMT mathematical reasoning benchmarks.

The Instruct variant targets low-latency inference, while the Thinking variant enables longer deliberation traces with explicit tool calls — a documented distinction that shapes how each is deployed in agentic workflows.

Access

Qwen3-Max-Instruct is available via Qwen Chat and through Alibaba Cloud Model Studio using an OpenAI-compatible API. The Thinking variant was in active training at the time of the official announcement and is being prepared for public release.