Alibaba's Qwen4-Max debuts: 2.4T MoE, 1M context, 62% cheaper per task than Qwen3.8-Max

AGI Hunt · wechat · 2026-11-16

Alibaba released Qwen4-Max, the first Qwen4 model: a new sparse MoE with 2.4T total parameters (71B active), native vision input, and a 1M-token context — priced below its predecessor, with cache costs cut 70%. It uses 24-48% fewer tokens than the last gen, dropping average Work360 task cost from $4.43 to $1.67. It tops DeepSWE v1.1, Agents' Last Exam v1.1 and a legal-agent bench (double Opus 5.5), though it trails Kimi K3.1 on Terminal-Bench 4.0. Demos span Blender/Unity/UE game building, a real auto-manufacturing pipeline (programming time -95%), an AI-designed coolant molecule validated in an automated lab, finance/legal/medical agent workflows, and 7×24 long-horizon delivery. Qwen4-Max itself stays closed; smaller Qwen4 models will be open-sourced.

Related event: Alibaba Releases Qwen4-Max with MoE Architecture, Native Vision, and 1M Token Context(15 posts)→

Original post →

More from Models

Models channel →