Alibaba’s Qwen4-Max Arrives as a 24/7 AI Teammate

Human Verified 0

Share to: , and

A glowing sphere of linked teal nodes hovers above a desk at dusk beside a small robot arm and a holographic chart

Alibaba launched Qwen4-Max on Monday, the first model of its new Qwen4 family, pitching it as a round-the-clock AI teammate for work, creativity and discovery. The flagship carries a 2.4-trillion-parameter backbone, native vision and a million-token context window, and is live on Alibaba’s cloud from today.

What the new model is

Qwen4-Max runs on a new sparse mixture-of-experts design, with 71 billion parameters active per token and 328 billion more in N-gram embedding layers. Alibaba rebuilt attention, residual connections and embeddings around Gated DeltaNet plus Qwen Sparse Attention for long sequences.

The model takes text, image and video input, and its context stretches to a million tokens, with up to 262,144 tokens of thinking plus final answer. It is available now on QwenCloud through the Token Plan or a pay-as-you-go API, and inside QwenWork, Qoder and QwenStudio.

Cheaper by the task

Alibaba’s headline claim is cost, not raw power. On its own Work360 evaluation of industry work, the average cost per task fell from $4.43 with Qwen3.8-Max to $1.67, a 62 percent cut, as token use dropped by roughly a quarter to a third.

API prices fell too: $1.50 per million input tokens and $4.50 per million output tokens, with implicit caching at $0.075, all below Qwen3.8-Max. The contrast with third-party hosts is sharp — one raised its Qwen3.7-Max prices by two-thirds earlier this month.

Built for long jobs

The launch leans hard on long-horizon work. Qwen4-Max ships a usable first version early and improves it in stages, a pattern Alibaba calls Live Delivery, and it is open-sourcing the plugin behind it so other builders can copy the approach.

Its boldest demos come from the lab. In a continuous 48-hour run on Deep Principle’s Mira platform, Qwen4-Max screened 2.7 million candidate molecules and directed an automated wet lab through four design loops to a physical sample of a new coolant material.

In drug discovery on DP Technology’s PharmMaster, it made 454 tool calls and 3,239 metric evaluations across 24 candidate molecules, ending with a compound whose estimated binding affinity is about 3.6 times the reference. Wet-lab validation is still pending for the candidate.

Our Read

The pitch is economics, not supremacy. Qwen4-Max is sold as the cheaper, steadier worker for long agentic jobs, and the 62 percent figure is the number Alibaba wants remembered. Treat it as directional: Work360 is the company’s own benchmark for this kind of work.

What is independently checkable so far is thinner. Qwen4-Max debuted at number six on LMArena’s Code Arena WebDev board with a preliminary 1,750 points, the top Chinese model there. Full third-party scores are still to come, and the efficiency claim rests on Alibaba’s own tests for now.

The closed flagship is also a statement of strategy. Max stays hosted-only, while Alibaba promises smaller open-weight Qwen4 models on the same lineage. That split — paid flagship up top, open models below — is now the pattern across the big labs.

Why It Matters

For buyers of AI, the release is a price signal. A flagship that undercuts its predecessor by more than half per task puts pressure on rival pricing, and the million-token window makes long jobs the new battleground rather than single answers.

For Alibaba, it is a statement of intent after months of teasing. Qwen4-Max arrives as the company builds QwenWork, Qoder and QwenStudio around the same models, turning a model release into a workspace play aimed at developers and professional teams.

Background

Qwen4-Max first appeared as a Token Plan preview on 11 November, after Alibaba teased the Qwen4 family with a riddle and no release date. A mystery model spotted inside Qoder days earlier was later identified in press reports as an early Qwen4-Max checkpoint.

Source: Qwen launch blog and QwenCloud docs

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.