Qwen Team Releases Qwen4-Max with 2.4T Parameters and 62% Cost Reduction

·2026.11.16 11:00

Key point

The 2.4T parameter model achieves a Work360 score of 52.69 and reduces average task costs from $4.43 to $1.67.

1 / 6

Details

  • Qwen Team released Qwen4-Max on 2026-08-03, featuring a 2.4T parameter backbone, native vision, and a 1 million token context window.
  • Average task costs dropped 62% to $1.67 on the Work360 benchmark, while token usage decreased by 24% to 49% depending on effort levels.
  • The model achieved a Work360 score of 52.69, surpassing Qwen3.8-Max's 47.74, with improved performance in coding, business, and long-horizon tasks.
  • Architecture utilizes GDN + Qwen Sparse Attention (QSA) and Gated Residual (GR) to compress history and stabilize training without loss spikes.

Performance and Efficiency

  • Token efficiency improved significantly, with Work360 token usage down 24% (Xhigh effort) and 35% (High effort).
  • Terminal-Bench 4.0 token usage decreased by approximately 48% and 49% for Xhigh and High effort settings, respectively.
  • Internal pre-training evaluations show Qwen4-Max-Base records the lowest BPB across five checkpoints, outperforming Qwen3.8-Max-Base by 5–13% with 25% fewer active parameters.

Industrial and Professional Applications

  • Manufacturing workflows reduced manual programming time from 475 minutes to 24 minutes, a 95% efficiency gain.
  • Legal due diligence generated an 86-page report with 14 specific legal issues identified from 198 working documents.
  • Critical care analysis integrated 1,433 medical images and 10,000+ character notes to propose management recommendations matching physician orders.

Scientific Discovery and Safety

  • Autonomous design of a novel coolant molecule completed in 48 hours with 4 iterations, validated by physical synthesis.
  • Drug discovery for NLRP3 inflammasome involved 454 tool calls and 3,239 computational evaluations, yielding a molecule with 3.6x binding affinity of the reference.
  • Security benchmarks show function-level secure code generation rising from 8% to 14.7%, with strong performance in avoiding injection and memory safety issues.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.