News

Alibaba Releases Qwen4-Max With a 1M-Token Window and 25% Lower API Prices

Alibaba made Qwen4-Max generally available on November 16 in Qwen Studio and through its API, at prices 25% below Qwen3.8-Max and with a one-million-token context window. The launch…

4 min read
Alibaba Releases Qwen4-Max With a 1M-Token Window and 25% Lower API Prices

Alibaba made its flagship multimodal model, Qwen4-Max, generally available on November 16, 2026. As reported by TechNode, the general rollout follows a preview period that began on November 11. The headline for anyone paying per token is the price: input and output are 25% cheaper than Qwen3.8-Max, and the model reads up to a million tokens of text, images and video.

What Is Actually New Under the Hood

According to the official launch announcement, Qwen4-Max features a sparse mixture-of-experts architecture totaling 2.4 trillion parameters. It activates 71 billion parameters per token alongside 328 billion N-gram embedding parameters, which Alibaba says is 25% fewer active parameters than Qwen3.8-Max-Base.

Sponsored

Tired of juggling ten tabs? ToolSuite bundles the AI workflow tools power users rely on — in one place.

Try ToolSuite Now

The model processes text, image, and video inputs, producing text across a 1,000,000-token context window. Deep thinking mode is enabled by default, supporting up to 262,144 combined tokens across reasoning steps and final output. For tooling, Alibaba showcased workflows integrating Blender, the MindOpt solver, and game engines including Unity, Unreal Engine, and Godot, and open-sourced a Live Delivery plugin that lets an agent hand over a first version of its work and revise it as you give feedback. These are Alibaba’s own demos, not independent tests.

Deployment Platforms and the Open-Weight Question

You can use the model in Qwen Studio (chat.qwen.ai), in QwenWork and in the Qoder coding tool. The announcement does not say what limits apply there or what consumer plans cost, so we can’t tell you what free users get.

Developers can access the API through QwenCloud and Alibaba Cloud Model Studio using the qwen4-max model ID. Notably, Qwen4-Max is a proprietary, closed model. While Alibaba announced that smaller open-weight models in the series will arrive soon, no specific release date was given. Until those arrive, Qwen4-Max runs only on Alibaba’s own services. Our Qwen tool page covers the wider model family.

Pricing Reductions and API Realities

The primary advantage of Qwen4-Max lies in published API pricing rather than vendor-measured benchmarks. While Alibaba’s blog claimed internal Work360 task costs dropped from $4.43 to $1.67, that is a figure from its own benchmark; the list prices are what you can count on. Base input and output costs fell by 25% compared to the previous generation, alongside substantial discounts on cached inputs.

Official rates published on the Alibaba Cloud pricing page detail the following comparisons:

MetricQwen3.8-MaxQwen4-Max
Input (per 1M tokens)$2.00$1.50
Output (per 1M tokens)$6.00$4.50
Cache read (per 1M tokens)Not listed$0.075

Enterprise users face several platform constraints. Model Studio does not support batch inference, fine-tuning, or dedicated instance deployment for Qwen4-Max. Additionally, while Alibaba offers 1,000,000 free tokens valid for 90 days, this tier is restricted exclusively to the Singapore region, offering zero free allowance in other deployment zones.

What Arab Users and Arabic Speakers Should Know

Alibaba’s announcement provides legal and clinical case studies in Chinese but includes no supported language inventory, leaving Arabic performance unconfirmed. This contrasts with the April 2025 predecessor release notes, which listed 119 languages and dialects, explicitly including Modern Standard Arabic alongside seven regional spoken dialects: Egyptian, Levantine, Najdi, Moroccan, Mesopotamian, Tunisian, and Ta’izzi-Adeni.

Model Studio lists regions in Beijing, Virginia, Frankfurt, Tokyo, Hong Kong, and Singapore, with none in the Middle East or Africa; the closest are Frankfurt and Singapore, and the free token allowance applies only in Singapore.

So don’t assume Arabic carried over. Run three quick checks in Qwen Studio first:

  • Dialect prompting: Ask the same question in Egyptian and in a Gulf dialect and see whether it understands local words and answers in clean Arabic.
  • Long document synthesis: Upload a long Arabic contract or PDF, ask for a clause-by-clause summary and check it against the original.
  • Visual OCR: Send a photo of a sign or invoice with Arabic text and ask for a word-for-word transcription.

What to Do Now: Who Should Switch

If you already pay for Qwen3.8-Max through the API, switching cuts your per-token price by a quarter; test it on your own Arabic tasks first. Teams that need very long context or video input have a cheaper option to try.

Conversely, teams with strict data residency mandates or fine-tuning requirements should wait. Those who want to self-host should wait for the promised open-weight models, and read our report on Qwen 3.8 models trying to reach Alibaba servers before running one locally. Teams preparing local infrastructure can explore our guide on running models locally.

Categories: News
Share:

Comments

0

No comments yet.