WangDou logo
WangDou

Alibaba releases Qwen 4 Max with 2.4 trillion parameters, tops China's agent leaderboard, but keeps the weights closed

2026-11-17·WangDou AI Express·Alibaba / Qwen / Large Language Models

Stack up 2.4 trillion parameters, top the Chinese leaderboard, then tell everyone: want the download? No.

Three key facts

Alibaba released Qwen 4 Max on November 16, its most powerful model yet and the first flagship built on its new Qwen 4 architecture. It has 2.4 trillion parameters, handles text, images and video, and accepts up to one million tokens of context. It uses a mixture-of-experts design that activates 71 billion parameters at a time, down from 95 billion in Qwen 3.8 Max, letting a much larger model spend less compute per token.

On Arena's agent leaderboard it is the highest-ranked Chinese model, ahead of Tencent, Moonshot AI and DeepSeek. On Agent Arena specifically it placed 14th among individual models and made Alibaba the fifth-ranked lab overall, posting a preliminary 7.46 percent score across 14,727 sessions, the best among Chinese entries. On Code Arena's web development ranking it debuted sixth globally at 1,750 points, 74 above its predecessor, and placed eighth in branding and reference-based design.

Through QwenCloud it costs $1.50 per million input tokens and $4.50 per million output tokens, 25 percent below the prior flagship's $2 and $6 rates, with cached reused input at $0.075 per million tokens. But the weights are closed; Alibaba has promised smaller open-weight Qwen 4 models later.

Wang Dou's Take

Start with the line everyone will skim past: closed weights. Qwen became the global developer community's stand-in for Chinese open source over the past two years, and the reason was never leaderboard scores. It was that you could download it, fine-tune it, and run it in your own racks. Now the flagship does not come that way. It comes as metered API calls.

That is not a retreat. It is a business. A 2.4-trillion-parameter model is not something most people can run even if they download it; the memory bill alone would buy a company. What disappears is the middle group: teams that wanted Qwen for private deployment, domain fine-tuning, and data that cannot leave the country. For them, Qwen 4 Max just went from our model to Alibaba Cloud's service. The promised smaller open-weight release is a door left slightly open so that group does not all leave.

Then look at the activation number: 710 billion out of 2.4 trillion. That is the real technical progress, because it means inference cost can be pushed down. Leaderboard positions are for other people to read. Activation size is what saves you money.

One more thing. A 7.46 percent agent score on 2.4 trillion parameters tells you that agent performance is not something you can just stack parameters at.

Source: Decrypt

Comments

Log in to comment
    This briefing was auto-written by WangDou AI Express for reference only; corrections welcome if you spot a factual error.
    指挥舱👽