Overview
Qwen4-Max is Alibaba's flagship model, released on 16 November 2026 as the first public entry point to the Qwen4 architecture. It carries 2.4 trillion total parameters but activates 71B per token, with a separate 328B pool of N-gram embedding parameters, native vision input and a 1M token context window.
What actually changed is not the score. Artificial Analysis puts its Intelligence Index at 43, nine points above Qwen3.8-Max at the xhigh setting and three behind Kimi K3.1 Pro. Nine points is an ordinary generational step. The price per task is what moved: $3.27, roughly a third of what Qwen3.8-Max costs per task, with 39 percent fewer output tokens. A model that gains a little and then cuts unit cost to a third means work you previously rationed can now run freely.
On capability, long-horizon agent work is what the vendor emphasises. It ranks first among Chinese models on Agent Arena, takes 1750 on the Code Arena WebDev board for sixth place and 1674 on Image-to-WebDev, also sixth, 35 points above Qwen3.8-Max. Vals Index accuracy is 63.20 percent at eleventh overall, again the highest for a Chinese model. The leaderboards measure different things, but they point the same way: the tuning targets sustained work rather than single-turn chat.
Availability needs to be stated plainly. Qwen3.8-Max shipped as open weights. Qwen4-Max is API only, at $1.50 input and $4.50 output per million tokens on Alibaba's first-party endpoint. Within two or three days OpenRouter, Novita AI and Vercel AI Gateway added it at the same prices. Teams that need local or private deployment have no weight release to work with as of 17 November; Unsloth only mentions that smaller open Qwen4 models are in preparation.
One detail that is easy to misread: the 2.4T on Qwen3.8-Max and the 2.4T here are not the same number. The older model activates far more per token, while this one holds activation at 71B and carries a separate embedding pool, so memory footprint and inference cost behave quite differently at the same headline size. Comparing the two by total parameters alone leads to the wrong conclusion.
Key Features
- 2.4T total parameters, 71B activated: A sparse MoE layout with a separate 328B N-gram embedding pool that offloads lookup-style recall, keeping inference cost well below a dense model of the same headline size.
- 1M token context: A whole repository, a hundred-page contract or a long meeting transcript can go in at once without slicing and summarising first.
- Native vision input: Images and text share one context, so screenshots, design files and receipts no longer need a separate vision model to describe them.
- Tuned for long-horizon agents: First among Chinese models on Agent Arena, 1750 on Code Arena WebDev and 1674 on Image-to-WebDev, with sustained execution named as the target workload.
- $3.27 per task: The vendor cites Artificial Analysis at roughly a third of Qwen3.8-Max per task, with 39 percent fewer output tokens, which compounds on long answers.
- Several entry points at launch: First-party API first, then OpenRouter, Novita AI and Vercel AI Gateway, all at $1.50 input and $4.50 output per million tokens.
Use Cases
- Question answering and rewriting across an entire codebase or long document set
- The controlling model in a long agent loop that decomposes work and calls tools
- Generating a working web page directly from a screenshot or design file
- Image-and-text understanding in Chinese business material such as receipts and contracts
- Bulk content production where cost per task is the binding constraint
Pros
- About $3.27 per task, low for a flagship of this generation
- 1M token context removes the need to slice long material
- Native vision avoids bolting on a separate description model
- 39 percent fewer output tokens than the previous generation
- Identical pricing on the first-party API and three aggregators, so switching does not reset the maths
- Chinese and English technical text sit in the same quality band
Pricing
The first-party API charges $1.50 input and $4.50 output per million tokens. Artificial Analysis estimates $3.27 per task, about a third of Qwen3.8-Max. OpenRouter, Novita AI and Vercel AI Gateway are live at the same prices. The open weights of Qwen3.8-Max do not apply here; Qwen4-Max is API only for now.
Summary
The score gain is ordinary; the value is in the cost curve. Nine index points come with a third of the per-task spend, plus 1M context and native vision, which makes long-material processing a routine rather than a special case. Teams that need open weights or private deployment cannot pick this generation yet.
Version History
- Qwen4-Max release (2026-11-16): Alibaba released Qwen4-Max, the Qwen4-generation flagship, with 2.4T total parameters, 71B activated per token, a 328B N-gram embedding pool, native vision and a 1M token context. Intelligence Index 43, nine points above Qwen3.8-Max, at $3.27 per task, roughly a third of the predecessor, with 39 percent fewer output tokens. Priced at $1.50 / $4.50 per million tokens, launched on the first-party API and followed by OpenRouter, Novita AI and Vercel AI Gateway.