Particle.news

Alibaba Launches Qwen4‑Max Flagship AI With MoE Design and Hosted‑Only Access

The hosted-first rollout signals a shift toward lower per-task costs under Alibaba’s control and raises the prospect of stronger customer reliance on its cloud services.

Overview

  • Alibaba formally launched Qwen4‑Max on Monday, November 16, 2026, and is offering access through its QwenCloud, QwenWork, Qoder and QwenStudio platforms with published API prices of $1.50 per million input tokens and $4.50 per million output tokens.
  • The model uses a mixture‑of‑experts (MoE) architecture with a 2.4 trillion‑parameter backbone, about 71 billion parameters activated per token, native vision support and a one‑million‑token context window to handle large codebases and long agent tasks.
  • Alibaba reports big efficiency gains versus Qwen3.8‑Max, saying Qwen4‑Max uses roughly 24–50% fewer tokens on internal tests and cut average cost per task on its Work360 benchmark by about 62% based on the company’s figures.
  • Qwen4‑Max is closed‑weight: Alibaba did not release the model weights and says smaller Qwen4 variants will be published as open weights later, a change from the partial openness it showed with Qwen3.8‑Max.
  • Independent evaluations give a mixed picture: Artificial Analysis scored Qwen4‑Max 43 and placed it near the middle of its Intelligence Index, while live agent leaderboards like Agent Arena ranked it below top rivals, highlighting a gap between vendor benchmarks and third‑party results.