Qwen Releases Qwen4-Max: 2.4T Parameters Put It Top of Agent Arena Among Chinese Models
The model releases are coming in a cluster this week, and Qwen has put out its own. On November 16, Qwen's official blog laid out the full architecture of Qwen4-Max: built on the new Qwen4 architecture, with a 2.4-trillion-parameter backbone, 71 billion parameters activated per token, plus a separate 328-billion-parameter N-gram Embedding, native vision support, and a context window stretched to one million tokens.
The same day, the Agent Arena leaderboard updated. With a +7.5% net gain, Qwen4-Max climbed to 14th overall and became the top-ranked Chinese model on the board. The organizers say the ranking draws on 14.7k real agent sessions.
What makes this worth reading is not just that the numbers got bigger. Two shifts are landing at once: models are heading toward "large and capable," and leaderboards are starting to measure them with real tasks rather than quiz answers.
Two: what the 2.4T number actually means
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.

The wording matters, or it is easy to misread.
2.4T is the backbone parameter count, not what runs on every call. What actually wakes up is the 71B activated per token — a sparse-activation (MoE) approach: the model holds a huge inventory, but a given token calls on only a small slice of it. A big warehouse, with only the needed lights switched on.
There is also a separate 328B N-gram Embedding, which is a different thing from the backbone and should not be folded into the same figure. Add native vision and a one-million-token context, and the positioning is clear: read very long material, understand images, and handle work that needs multi-step reasoning.
The official release pairs the architecture details with cost comparisons and industry case studies — the message is that you should not just admire the parameters, you should be able to work out what using it costs.
Three: a Chinese first, with a caveat
Qwen4-Max sits at 14th overall on Agent Arena and is the top Chinese model on the board. As a result, Alibaba became the highest-ranked Chinese lab and fifth overall, behind Anthropic, Google, OpenAI, and Meta. That phrasing matters: this is a Chinese first, not a world first — four labs remain ahead on the overall board.
What gives the ranking weight is its method. It draws on 14.7k real agent sessions broken into four signals, measuring whether a model can carry real work through rather than answer quiz questions. A separate report put Qwen4-Max 9 points above the previous Qwen3.8 Max on the Intelligence Index, at lower cost. Smarter and cheaper at once is where this round of competition is actually moving.
Four: three things worth keeping

One, split the parameters apart when you read a sparse model. Backbone size and activated size are two different things. 2.4T is the warehouse; 71B is the per-call cost. Conflating them makes you over- or underestimate the real expense.
Two, check how a leaderboard ranks before you trust it. A ranking based on real agent sessions says more about "can it do the work" than a quiz score does — but it is still one lens, not a final verdict.
Three, the more models move to the cloud, the more the local machine matters. Running inference in the cloud is fine. But dispatching the jobs, storing the files, and keeping the memory need somewhere that is yours.
Models like Qwen4-Max are bound to run in the cloud. The machine that calls them still needs to stay awake. For ordinary users who want to put this to work, a natural companion is KAIHE AIBOX(铠盒AIBOX), a computer built specifically for agents and your local private AI assistant. OpenClaw and Hermes come installed from the factory, so it is ready on standby the moment it is plugged in and online — no rebuilding the environment again and again. Inference goes to the cloud, while files, long-term memory, and call records stay local.
Models can be swapped out every time a new one ships. The machine that runs errands for you should preferably be your own.