Alibaba launches Qwen4-Max: 2.4T-param MoE with native vision, cuts per-task cost 58%
千问大模型 · wechat · 2026-11-16
Qwen officially released Qwen4-Max, built on a new sparse MoE architecture: 2.4T total parameters with 71B activated per token plus 328B N-gram Embedding parameters, native vision, and a 1M-token context window. It markedly improves over Qwen3.8-Max and ranks near the top of the CodeArena WebDev blind leaderboard; smaller Qwen4 models will be open-sourced progressively.
Architecture & cost: upgrades across attention, residual and embedding — GDN + QwenSparseAttention selects context at block granularity, GatedResidual runs four parallel residual branches, N-gram Embedding adds capacity cheaply; pretraining had no loss spikes, with BPB down 5–13% and 25% fewer activated parameters vs the 3.8 generation. Work360 token consumption dropped 24%/35% (Xhigh/High) and 48%/49% on Terminal-Bench4.0; after across-the-board API price cuts, average cost per Work360 task fell from ¥26.58 to ¥11.16 (−58%), with cache hits as low as ¥0.5 per million tokens.
Capabilities: Work360 industry benchmark rose from 47.74 to 52.69; it drives Blender/Unity/Unreal/Godot for games and 3D, and in a real automotive production scenario cut per-part machining programming labor from 475 to 24 minutes (−95%). It can call 90 MCP tools for auditable financial research, produce an 86-page IPO legal due-diligence report from 198 messy source files, and read a full critical-care record including 1,433 images in one session to reach doctor-aligned decisions. Long-horizon agents get LiveDelivery continuous shipping (with the open-sourced Qwen-Live-Delivery-Plugin); the model already optimizes Alibaba Cloud's own GPU/CPU scheduling. In AI4Science, it autonomously designed and physically synthesized a new coolant molecule over 48 hours and optimized Alzheimer's drug candidates against NLRP3 (3.6× estimated binding affinity vs the patent starting point).
More from Models
- User blasts Claude subscription model: paid tokens vanish if you take a break — RileyRalmuto · 2026-11-16
- Qwen4-Max lands #6 in Image-to-WebDev Arena at 1674 pts, up 35 over predecessor — Alibaba_Qwen · 2026-11-16
- Users can't access any Gemini model in Google's Antigravity, forced to use Claude — rez0__ · 2026-11-16
- Alibaba's Qwen team releases flagship Qwen4-Max, pitched as 'your new teammate' — byte-style · 2026-11-16
- Explainer video: how tokens give LLMs their language and agentic superpowers — Substantial-Farmer31 · 2026-11-16
- Qwen4 Max scores 43 on AA Intelligence Index, 67% cheaper per task than predecessor — ArtificialAnlys · 2026-11-16