Alibaba Unveils Qwen4-Max: 2.4T Parameters, 58% Lower Cost per Task
APPSO · wechat · 2026-11-17
APPSO got early hands-on time with Alibaba's new flagship Qwen4-Max — the first flagship since Liu Dayiheng took charge of Qwen.
Key specs and results
- Brand-new architecture: 2.4 trillion total parameters with only 71 billion activated per token (about a quarter less than the previous generation), native vision capabilities, and support for a 1-million-token context window.
- On the Terminal-Bench 4.0 coding benchmark, token consumption was nearly halved while the score rose from 38.9 to 48.5; API input pricing dropped to 10 yuan per million tokens, with cache hits as low as 0.5 yuan.
- Scored 1750 on CodeArena's blind web development evaluation, ranking 6th globally — the top Chinese model on the leaderboard. It topped long-horizon task evaluations ahead of Claude Opus 5.5 and Kimi K3.1, with per-task costs nearly 60% lower than its predecessor.
Hands-on highlights
- Asked the model to build interactive teardown demos of the iPhone 17 Pro and Cybertruck from scratch using Three.js — geometry, textures, and part placement were all generated via code, showing excellent 3D spatial understanding.
- Cross-software collaboration: designing an enclosure in OpenSCAD, running circuits in ngspice, and creating assembly animations in Blender, with all three matching perfectly in physical dimensions.
- Alibaba named this "ship a working version first, then iterate on feedback" workflow "Live Delivery" — a sample physics education website went through 14 delivered versions.
Behind the scenes
- Continuing the recursive self-improvement (RSI) vision from the Apsara Conference: Qwen4-Max optimized its own scheduling system, cutting annotation task scheduling costs by 10.42% and making offline CPU rescheduling 14.61x faster.
- Upcoming Qwen4.5/Qwen5 will reach 5–10 trillion parameters; Qwen4-Max won't be open-sourced for now, as smaller models are released gradually.
More from coding & agent
- Agent Substrate: open-source zero-trust runtime suspends idle agents, fits 10x more per hardware — Saboo_Shubham_ · 2026-11-17
- ChatGPT Wallet solves secure payment, not buying the right thing — krishnan · 2026-11-17
- Grok bot phones a store and secures a $1,511 refund for its user — elonmusk · 2026-11-17
- Practical tips to cut agent memory costs on OpenAI: cache TTL, Flex, fewer calls — gabrielchua · 2026-11-17
- 10 open-source repos to build an entire AI content factory solo — aftahi_ai · 2026-11-17
- What 4 calibration papers do (and don't) say about gating agent tool calls on model confidence — Xjjjjyn · 2026-11-17