Alibaba's Qwen4-Max Is a Flagship You Can't Download
By @sharedot · · 6 pages
- AI Frontier
- Open Weights
- Qwen4 Max
- Alibaba
Alibaba's Qwen4-Max tops Chinese agent leaderboards, but the flagship ships closed-weight — a twist for the world's most-downloaded open-model family.
The release, and the catch nobody expected
Alibaba launched Qwen4-Max on Monday, its most capable model yet and the first flagship on the new Qwen 4 architecture, with 2.4 trillion total parameters, native vision capabilities, and a one-million-token context window.
The company says it excels at complex tasks across work, creativity, and research. But Decrypt reports the twist for the developer community: the model is closed-weight.
Developers can use it via API or Alibaba's apps, but they cannot download its internal parameters, modify it, or run it on their own hardware — an unusual posture for a family built on open distribution.
A reversal for the open-weight ecosystem
The closed flagship is a striking reversal because, according to Bloomberg News as cited by PYMNTS, the Qwen series is so far the world's most-downloaded open-weight model family, pushed by Alibaba CEO Eddie Wu.
Decrypt adds that an August Hugging Face report counted 151,448 Qwen-based derivative models — about 4.7 times the number built on Meta's Llama family. Alibaba did open the weights of Qwen 3.8 Max earlier this year, so the community expected the same of the new flagship.
Instead, Decrypt reports Alibaba only promises that smaller Qwen 4 models will ship with open weights, with no sizes or release dates announced.
What the benchmarks actually show
Qwen4-Max is the highest-ranked Chinese model on Arena's agent leaderboard, per Decrypt, posting a preliminary 7.46% score across 14,727 sessions and making Alibaba the fifth-ranked lab overall — ahead of Tencent, Moonshot AI and DeepSeek.
On Code Arena's web-development ranking it debuted sixth globally at 1,750 points, up 74 from its predecessor. Alibaba says the model outperforms its predecessor Qwen3.8-Max while using 35% fewer tokens in high-effort reasoning and cutting average cost per task by roughly 60%.
Decrypt reports the mixture-of-experts design activates 71 billion parameters at a time, versus 95 billion for Qwen 3.8 Max.
Pricing, rivals, and what comes next
Through QwenCloud, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens — a 25% cut from the prior flagship's $2 and $6 rates, per Decrypt, with cached reused input at $0.075 per million.
Alibaba says the new model outperformed rivals including Anthropic's Claude Fable 5.1, Moonshot's Kimi K3.1 Pro and Muse Wonder Full. Decrypt cautions that American labs still hold the top spots in Arena's overall agent and web-development tables, though at considerably higher prices.
Builders watching the open ecosystem will want the promised smaller open-weight Qwen 4 models — and any announcement of when they land.