qwen 4 max: Alibaba's 2.4T Flagship Tops China's Agent Charts, No Weights
The qwen 4 max launch gives Alibaba a 2.4-trillion-parameter flagship with 1M-token context and cheaper API pricing, ranked sixth on Code Arena, but its weights stay closed.

- Qwen4-Max uses a mixture-of-experts design that activates 71 billion parameters per token, 25% fewer than Qwen3.8-Max, per Alibaba.
- API pricing on QwenCloud is $1.50 per million input tokens and $4.50 per million output tokens, about 25% cheaper than the previous flagship.
- It is closed-weight; Alibaba says smaller open-weight Qwen4 models are coming soon, without dates or sizes.
in this block
The qwen 4 max launch puts Alibaba back at the front of China's AI race. Alibaba released Qwen4-Max on Monday, a 2.4-trillion-parameter flagship with a one-million-token context window and native image and video understanding, and says it ranks sixth on Code Arena, according to Alibaba Cloud and Decrypt. The catch for open-source fans: you cannot download it.
What actually happened
Alibaba Cloud's press release, dated Nov 16 in Hangzhou, calls Qwen4-Max "the most capable model in the Qwen family to date." The company says it leads in coding, agentic tasks and visual reasoning, and that it outperforms Qwen3.8-Max on Alibaba's internal Work360 benchmark while using 35% fewer tokens at high reasoning effort and cutting average cost per task by around 60%.
Decrypt's Jose Antonio Lanz dug into the leaderboards. On Arena's agent leaderboard, the qwen 4 max model is the highest-ranked Chinese entry, ahead of Tencent, Moonshot AI and DeepSeek in that test, with a preliminary 7.46% score across 14,727 sessions. On Code Arena's web-development ranking it debuted sixth globally at 1,750 points, up 74 from its predecessor.
Bloomberg, via Yahoo Finance, reported that Alibaba claims coding performance ahead of Anthropic's Claude Fable 5.1 and Moonshot's Kimi K3.1 Pro. That is Alibaba's claim; independent testing will take time.
Alibaba's pitch: frontier-ish performance, lower token bills, and no weights for you.
The specs that matter
Here is the short version from Alibaba and Decrypt:
- 2.4 trillion total parameters, 71 billion active per token.
- Up to one million tokens of context, enough for big codebases or document stacks.
- Text, image and video input; outputs are text and code, not a new image generator.
- Cached input as low as $0.075 per million tokens, which helps repetitive agent workflows.
Alibaba Cloud's model release log lists qwen4-max as available for international deployment from Nov 16, after a preview version appeared on Nov 11.
Context length is the other headline. A one-million-token window lets developers feed in long codebases, contracts or research packs in one go, which is exactly what coding agents and document assistants need. Alibaba also says the model supports tool calling, web search and structured outputs, and is compatible with OpenAI and Anthropic style APIs, according to its model release notes.
Why the closed weights sting
Qwen became huge because of open models. Decrypt cites a Hugging Face report from August counting 151,448 Qwen-based derivative models, about 4.7 times the number built on Meta's Llama family. Alibaba opened the weights of Qwen3.8-Max earlier this year, so keeping the qwen 4 max weights closed is a notable shift.
It also lands in a crowded Chinese field. Our kimi k3 1 pro piece covered Moonshot's latest push, and our deepseek v4 1 flash 1017 story showed DeepSeek competing hard on price. American labs still hold the top overall spots on Arena's agent and web-dev tables, Decrypt notes, but they also charge more.
How it compares on price
The pricing is a big part of the pitch. Decrypt says Qwen4-Max costs about 25% less than Qwen3.8-Max's $2 and $6 per million tokens, and Alibaba claims tasks run about 60% cheaper on average because the new model uses fewer tokens. If those numbers hold in real use, agent workloads that burn millions of tokens a day get noticeably cheaper.
The trade-off is control. With a closed model, you depend on Alibaba's API, its regions and its policies. Teams that need to self-host will have to wait for the smaller open-weight Qwen4 releases Alibaba has promised.
What to do as a reader
If you build with LLMs, the qwen 4 max release is worth a test run, especially for agent and front-end work where token costs add up.
- Benchmark it on your own tasks; vendor demos and early leaderboards are not the same as your workload.
- Check data residency and compliance rules before sending sensitive data to any hosted API.
- Watch for the promised open-weight Qwen4 models if you need local or private deployment.
- Ignore "Qwen" tokens on DEXs; there is no sign Alibaba launched any crypto token tied to this model.
- Read Alibaba Cloud's announcement for the official claims.
Cheaper smart tokens are good for builders. Just keep your own evals honest. Wagmi.
Not financial advice. DYOR, ser.