Qwen4‑Max employs a sparse mixture‑of‑experts design, activating 71 billion parameters per token and supporting multimodal text, image and video inputs within a 1 million‑token context window.
Alibaba has unveiled Qwen4‑Max, a new multimodal mixture‑of‑experts (MoE) model that boasts a staggering 2.4 trillion parameters and can process text, images and video within a one‑million‑token context window.
Key technical specifications
Qwen4‑Max uses a sparse MoE architecture that activates roughly 71 billion parameters for each token, allowing the model to scale efficiently while maintaining high performance across modalities.
- 2.4 trillion total parameters
- 71 billion active parameters per token
- Supports text, image and video inputs
- 1 million‑token context length
Multimodal capabilities
The model can ingest long‑form documents, high‑resolution images and continuous video streams, enabling applications such as extended document summarisation, visual question answering and video‑to‑text generation.
Potential use cases
Enterprises may leverage Qwen4‑Max for tasks that require deep contextual understanding, including legal contract analysis, e‑commerce product recommendation with visual cues, and automated video captioning for media platforms.
For readers interested in broader AI developments, explore the AI coverage on uxc.news.