- Scores 43 on the Intelligence Index, up 9 points from Qwen3.8 Max with gains on every included evaluation
- Cost per task is $3.27, about one-third of the predecessor and ~11% below Kimi K3.1 Pro
- Uses 39% fewer output tokens, with cache hits priced at $0.075
AI CURATED HOT STORY · 11/16/2026
Alibaba Unveils 2.4T-Parameter Qwen4-Max with 60% Cost Reduction
Alibaba's Tongyi Qianwen team has launched Qwen4-Max, a flagship model with 2.4 trillion parameters, sparse MoE architecture, native vision understanding, and a 1 million token context window. The company claims over 60% lower task cost at equivalent performance, now available via QwenCloud API.
The model introduces architectural innovations like Gated DeltaNet and Qwen Sparse Attention to reduce long-sequence overhead. Third-party evaluations show a 9-point intelligence index gain, with per-task cost at $3.27, about one-third of its predecessor. On Agent Arena, it jumped from 39th to 14th, becoming the highest-ranked Chinese model, and rose to 6th in coding arena.
Qwen4-Max marks a key breakthrough in efficient large models, potentially reshaping the economics of enterprise AI and intensifying global competition. Smaller open-source variants are expected to further expand the developer ecosystem.
Hourly heat curve
2 fully observed hoursCoverage timeline
Total 5 reports · Latest first- Artificial Analysis (@ArtificialAnlys)T1·84 pts
- Arena.ai (@arena)T1·72 pts
- Qwen4-Max ranks No. 6 on Code Arena WebDev at 1750 points, +74 over Qwen3.8-Max and the top Chinese model there
- Agent Arena rank climbs from No. 39 to No. 14 overall, with a 7.5% net gain across 14.7k real sessions
- Alibaba becomes the No. 1 Chinese lab and No. 5 lab overall on Agent Arena, behind Anthropic, Google, OpenAI and Meta
- Arena.ai (@arena)T1·78 pts
- Qwen4-Max climbed from #39 to #14 overall on Agent Arena, becoming the top-ranked Chinese model
- Delivered a 7.5% net gain across 14.7k real-world agentic sessions, with the largest jumps in user satisfaction and steerability
- Implicit caching pricing dropped 70%; API is live on QwenCloud and available in Qwen Studio, QwenWork and Qoder
- Hacker News AIT2·88 pts
- Built on the new Qwen4 architecture with a 2.4-trillion-parameter backbone and native vision
- Context window reaches one million tokens, targeting long-horizon tasks and professional tool use
- Qwen claims broad gains over Qwen3.8-Max across coding, work and long-horizon benchmarks
- IT之家 · 智能时代T2·88 pts
- 2.4T-parameter sparse MoE with native vision and a 1M-token context window
- Three of every four layers use Gated DeltaNet to compress history into fixed states; the fourth applies block-level Qwen Sparse Attention to cut long-sequence cost
- Token consumption at Xhigh/High reasoning effort drops ~24% versus the prior generation, with task cost down over 60%
- Available now on QwenCloud API, with smaller same-lineage models to be open-sourced