Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context

Impact 8 MarkTechPost

Summary

  • Alibaba has released Qwen4-Max, a proprietary sparse Mixture-of-Experts model featuring a 2.4-trillion-parameter backbone and 71 billion active parameters per token.
  • The model supports a 1-million-token context window, accepts text, image, and video inputs, and is available exclusively via hosted API platforms such as QwenCloud.
  • Architectural updates include Qwen Sparse Attention, Gated Residual connections, and 328 billion N-gram embedding parameters to improve context processing and training stability.
  • Alibaba reported a 62% reduction in average task cost alongside lower API pricing, with input tokens priced at $1.50 per million and output tokens at $4.50 per million.
  • While scoring competitively on agentic and legal tasks like HLAB-Vals-1.1, Qwen4-Max trails some frontier models on specific programming and vision benchmarks.

Why it matters

Qwen4-Max showcases significant architectural adjustments in sparse MoE scaling and long-context processing while reducing inference costs, providing enterprise developers a competitive hosted alternative for agentic workloads.

Who should care