Radar Telegram alerts Open Radar

Alibaba launched its flagship model Qwen4-Max with an original vision and the context of 1M token.

תרגם לעברית

Метаверсище и ИИще · · Source

🇨🇳 China · Map

Alibaba launched its flagship model Qwen4-Max with an original vision and the context of 1M token.

Alibaba launched its flagship model Qwen4-Max with an original vision and the context of 1M token.

Qwen4-Max is built on a new MoE architecture with native multimodal vision and a 1 million token context window. The model is already available via API on QwenCloud, as well as Qwen Studio, QwenWork and Qoder.

The main focus is on agent work efficiency and code generation: Qwen4-Max outperforms Qwen3.8-Max while spending up to 50% less tokens, which reduces the average cost of completing a task by more than 60%. In the independent Agent Arena rating, the model rose for the first time with a +7.5% net increase (14.7 thousand real tasks), and in Code Arena: WebDev it rose to 6th place (+74 points for Qwen3.8-Max).

At the same time, the Qwen team confirmed the upcoming release of open scales for junior models of the Qwen4 line - in particular, the compact Qwen4-27B will be able to work on systems with 16 GB of memory.

In summary:

A new MoE architecture with an original vision and context of 1M token.

API rates: $1.5/M input tokens, $4.5/M output and $0.075/M cache (-70% to Qwen3.8-Max).

Top 6 in the Code: WebDev arena (+74 points) and number 1 among Chinese models in the Agent Arena.

An open release of junior weights from the Qwen4 family is expected in the near future.

Sources:

The official Qwen4-Max blog.

QwenCloud API.

Agent Arena rating.

More on Radar

Want to see every report in real time?

Hundreds of sources worldwide, translated into your language, with alerts for what matters to you.

Sign up free
Share on WhatsApp Get alerts on Telegram