Skip to content

Qwen4-Max: Your New Teammate Has Arrived

Qwen

The MoE model features native vision, a one-million-token context window, 71B activated parameters per token, and is available via QwenCloud API and Token Plans.

Key points

  • Architecture incorporates 2.4T backbone parameters, 71B active parameters per token, and 328B N-gram embedding parameters.
  • Combines Gated DeltaNet layers with Qwen Sparse Attention to compress history into fixed-size states for long sequences.
  • Reduces Work360 task token usage by 24% to 35% and cuts average cost per task by 62% compared to Qwen3.8-Max.
  • Includes an open-source Qwen-Live-Delivery-Plugin for staged execution and real-time user steering of long-running agents.
  • Alibaba plans to release smaller models in the Qwen4 family as open source.

Why it matters

Developers and enterprises gain access to a larger multimodal MoE model optimized for long-horizon agent workflows, tool use, and complex domain tasks at reduced API costs.

Sources