Overview
- Qwen4‑Max, which Alibaba launched on Monday, November 16, 2026, is built on a 2.4‑trillion‑parameter sparse mixture‑of‑experts backbone, includes native vision capabilities, and supports a one‑million‑token context window.
- Alibaba says the model activates about 71 billion parameters per token and is materially more efficient than Qwen3.8‑Max, using roughly 48–50% fewer tokens and cutting average task costs by more than 60% in the company’s tests.
- The company set aggressive hosted pricing at $1.50 per million input tokens and $4.50 per million output tokens, and it has a Token Plan preview that opened on November 11 and runs through November 25, 2026.
- Alibaba published vendor benchmarks showing Qwen4‑Max leading on some coding tests (DeepSWE v1.1 score 75.7), but independent live‑agent rankings such as Agent Arena report mixed results and place Qwen4‑Max well below top competitors on real‑world task success.
- Alibaba kept the flagship weights closed while promising smaller open‑weight Qwen4 models later, a move that may lower run costs for cloud customers but reduce self‑hosting and increase vendor lock‑in as the company expands its cloud, tooling and chip ecosystem.