Alibaba launches Qwen4-Max: MoE, native vision, 1M context, tasks cost 60% less

OfirPress · x · 2026-11-18

Alibaba's Qwen team released Qwen4-Max, built on a new MoE architecture with native vision and a 1M-token context window. Internal evals claim 50% fewer tokens per task vs Qwen3.8-Max and over 60% lower average cost per task, with pricing at $1.5/M input, $4.5/M output and $0.075/M implicit-cached tokens. Separately, researcher KLieret noted Qwen's progress on ProgramBench, while stressing the main benchmark metric remains <20% for all models.

Related event: Alibaba launches Qwen4-Max with new MoE architecture and 1M context, topping Chinese models across arenas(64 posts)→

Original post →

More from Models

Models channel →