RU

Alibaba ships Qwen4-Max: 2.4 trillion parameters and a 60% price cut per task

Published: 2026-11-17 · Author: AI Release · @ai_release1
Alibaba ships Qwen4-Max: 2.4 trillion parameters and a 60% price cut per task

⚡ The gist

🔍 What happened

On November 17, Alibaba opened up general access to Qwen4-Max. The model runs on the company's cloud platform: via API, QwenCloud token plans, and inside QwenWork and QwenStudio. It's the direct successor to Qwen3.8-Max — with the same total parameter count of 2.4 trillion.

The mechanics, in numbers. The architecture is sparse MoE: 71 billion of the 2.4 trillion parameters spin up to process each token. The previous version had 25% more active ones. At high reasoning settings, the model burns 35% fewer tokens, and the average cost per task drops by roughly 60%. Context has grown to 1 million tokens, and Qwen4-Max parses images on its own, no separate vision module needed.

Alibaba backed the release with benchmarks. In Code Arena, the model ranks sixth. On its own Work360 benchmark — 721 professional domains — it beat its predecessor. In a demo with auto parts, the agent caught errors in a project on its own and cut the time to write a processing program from 475 minutes to 24. Offline task replanning on CPU got more than 14x faster, and model-call costs during labeling dropped by 9% on average.

There are research runs, too. In screening coolants for chips, an agent ran for 48 hours straight and completed four cycles — from molecular design to results analysis. One candidate was later validated in a real experiment, confirming its cooling properties. On the Alzheimer's front, the agent made 454 tool calls, reviewed 3,239 metrics, and produced 24 candidate molecules — and ZDNet Korea is upfront that this is search, not proven therapeutic effect.

🧠 Our take

The bet on cheap inference is obvious at a glance. Alibaba isn't chasing absolute benchmark records — it's selling economics: the same 2.4 trillion on paper, fewer active per token, fewer tokens per task. Our read: this is the answer to the fact that in 2026, the market isn't complaining about quality — it's complaining about the API bill.

What it changes. At $0.075 per 1 million cached input tokens, long agentic loops stop being a luxury. The demo with a 48-hour run and 454 tool calls is exactly about that: before, such experiments hit a budget wall, not a model-intelligence wall. For a team running an agent over documentation or logs, a 60% difference per task is the difference between a pilot and production.

Who wins. Alibaba — it's walking into enterprise pipelines through pricing and through QwenWork. Developers win too, as long as the price war lasts. The losers, it seems, are those selling "smarts" without economics: when a competitor delivers comparable results at a third of the price, benchmark questions become a distant second. But a 25% cut in active parameters can also come from aggressive routing. Our take: without external measurements on someone else's tasks, the "minus 60%" figure remains a vendor promise.

⚠️ What we don't know

The biggest weakness is the source of the numbers. Work360 is Alibaba's own invention, and sixth place in Code Arena says nothing about how far the model trails the top 5, or in which languages. It's unclear how the MoE routing behaves on non-English text and on messy corporate documents. The $0.075 price is cached input only — no output pricing, no cache misses. And one more thing to keep in mind: in October, Cohere publicly raised safety concerns about Qwen, so distrust of Chinese models hasn't gone anywhere. Verify it yourself — on your own tasks and your own bills.

💬 Worth debating

  • What matters more for agentic workflows: absolute quality on third-party benchmarks, or the cost per task in your own pipeline?
  • Should you wait for open-weight Qwen4 versions, or build production on a closed API that could get more expensive at any moment?
  • 24 candidate molecules in a single run — a breakthrough in automated science, or a slick demo nobody has reproduced yet?

Forecast: within the quarter, Alibaba will release at least one open model in the Qwen4 series — smaller than Max in parameters, with a focus on local deployment.

Source: ZDNet Korea

🤖 AI summary
  • Date: 2026-11-17
  • Source: zdnet.co.kr
  • Tags: Alibaba, Qwen4-Max, MoE, LLM, QwenCloud, AI, Qwen, neural networks
  • Key facts:
    • Alibaba has rolled out Qwen4-Max — a MoE model with 2.4 trillion parameters, of which 71 billion fire per token.
    • Available via QwenCloud through API and token plans, plus inside QwenWork and QwenStudio.
    • Context window goes up to 1 million tokens, with native vision; cached input costs $0.075 per 1 million tokens.
    • Smaller open-weight models in the Qwen4 series are promised later — no dates, no sizes.
    • On November 17, Alibaba opened up general access to Qwen4-Max.
    • The model runs on the company's cloud platform: via API, QwenCloud token plans, and inside QwenWork and QwenStudio.
    • It's the direct successor to Qwen3.8-Max — with the same total parameter count of 2.4 trillion.
    • The mechanics, in numbers.
#Alibaba#Qwen4-Max#MoE#LLM#QwenCloud#AI#Qwen#нейросети
← PreviousTensorWave to hand OpenAI 457 MW on AMD MI455X: deliveries start in 2027

Source: zdnet.co.kr · post in Telegram