FULL STORY

Qwen4-Max Launches: New MoE Architecture Meets First Benchmarks

Alibaba's Qwen team launched Qwen4-Max on Nov 16 with a new MoE architecture, native vision and a 1M-token context. While internal benchmarks claim ~50% token savings, Artificial Analysis scored it 43 on its Intelligence Index—cheaper to run but lower-ranked.

2026-11-16 ~ 2026-11-16 · 3 episodes · 24 posts

Episode 1 · Alibaba Releases Qwen4-Max with MoE Architecture, Native Vision, and 1M Token Context (2026-11-16, 15 posts)

Alibaba's Qwen team released its flagship model Qwen4-Max on November 16, positioning it as an "AI teammate" for work, creation, and discovery. Key takeaways: the model is built on an all-new MoE architecture, supports native visual input and a 1 million token context window, and is available via API only—officials confirmed smaller Qwen4-series models will soon be released with open weights. Notably, the team backed up claims about efficiency, cost, and professional work capabilities with extensive quantitative data, and shipped a new Work360 benchmark alongside.

Confirmed

  • Architecture and capabilities: all-new MoE architecture, native visual input, 1M token context window (m1, m7)
  • Token efficiency: official internal tests show token usage down roughly 24% (Xhigh tier) and 35% (High tier) versus Qwen3.8-Max, with overall savings of up to 49%–50%; cost per task drops to $1.67, a reduction of over 60% (m2, m8)
  • Work360 benchmark: covers 26 top-level domains, 182 second-level subdomains, and 721 third-level specialty areas; Qwen4-Max scores 52.69, above Qwen3.8-Max (m5)
  • Official demos: 3D design (from blockout to cinematic shots, m3), manufacturing (from sketch to real industrial production, m4), and research (autonomously designing and synthesizing a novel coolant material, m6)
  • Release format: API only; according to an announcement relayed by @ResearchCrafty1804, smaller Qwen4-series models will soon be open-weight (m7)

Why it matters

  • The combination of million-token context and native vision, plus dramatic cost reductions, targets professional workflows involving long documents and multimodal input—putting it in direct competition with rival flagships.
  • Rather than chasing general leaderboards, the team built its own layered Work360 work benchmark and rolled out a dense series of industry demos (3D, manufacturing, research), signaling a strategy focused on real-world work scenarios.
  • The tiered open-source roadmap (Max as closed API, smaller models open-weight) leaves room for the community to follow up.

Episode 2 · Qwen4-Max Benchmarks Leak: 50% Fewer Tokens, 60% Cheaper (2026-11-16, 3 posts)

Leaked Qwen4-Max benchmark charts claim a new MoE architecture beats Qwen3.8-Max with about 50% fewer tokens and over 60% lower per-task cost, with full comparisons against Opus5.5, GPT6.1 and Kimi K3.1, though some users remain skeptical.

Episode 3 · Artificial Analysis: Qwen4 Max Scores 43 on Intelligence Index, Cuts Task Cost 67% (2026-11-16, 6 posts)

Artificial Analysis benchmarks Alibaba's Qwen4 Max at 43 on the Intelligence Index, up 9 points over Qwen3.8 Max, with task cost down 67% to $3.27. Knowledge work beats Kimi K3.1 Pro, but coding lags and its 38/81 rank disappoints some observers.