Skip to content
D-CSIL™

AI News · 2026-11-16 · 7:04 AM CT

Alibaba's Qwen4-Max: same backbone, fewer active parameters, lower prices

Paul Derrington, Jr.

By Paul Derrington, Jr.

CEO | Chief Engineer | Research Engineer | Editor, D-CSIL

4 min read

Follow @PaulDerrington on X

Key Takeaways

Alibaba released Qwen4-Max on November 16, five days after a Token Plan-only preview. It keeps the 2.4-trillion-parameter backbone of Qwen3.8-Max but activates about 71 billion parameters per token instead of about 95 billion. Alibaba says fewer tokens per task and lower API prices cut the average cost per task on its own Work360 evaluation from $4.43 to $1.67. The API is live on QwenCloud at $1.50 per million input tokens and $4.50 per million output tokens, with a 1-million-token context window and native vision.

Qwen4-Max cover image from Alibaba's launch post for the model.
Photo: Qwen Team / Alibaba

Old vs. new: what stayed the same

The backbone did not grow. Both Qwen4-Max and its predecessor Qwen3.8-Max, unveiled in August, sit on a 2.4-trillion-parameter mixture-of-experts backbone, according to Alibaba's launch post. What changed is how much of it runs on each token: Qwen4-Max activates about 71 billion parameters per token, down from about 95 billion — about 25% fewer — plus an additional 328 billion N-gram Embedding parameters that add capacity with little extra compute.

Around that smaller active footprint, Alibaba rebuilt the architecture: Gated DeltaNet layers compress history into fixed-size states for its sparse-attention layers, four gated residual branches handle information flow, and context-based N-gram lookups sit at the second layer. The company says training at 2.4-trillion-parameter scale ran with no loss spikes, and that the base model scored 5–13% lower bits per byte than Qwen3.8-Max-Base across its in-house knowledge, code, QA and STEM evaluations while using those fewer active parameters. Inputs are text, images and video, with text out, and the context window is 1 million tokens.

The efficiency numbers, side by side

The headline claim is cost per task, not cost per token. On Work360 — Alibaba's own benchmark of challenging industry tasks across 721 specialized fields in 26 domains — Qwen4-Max scored 52.69, up from 47.74 for Qwen3.8-Max, while using about 35% fewer tokens at High reasoning effort and about 24% fewer at Xhigh. On Terminal-Bench 4.0, Alibaba says token use fell by about 48–49%. Combined with lower list prices, the average Work360 task cost fell from $4.43 to $1.67, a 62% reduction in Alibaba's evaluation, per the Qwen team's launch post; Alibaba's press release rounds it to around 60%.

List prices moved the same direction: $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6 for Qwen3.8-Max, with implicitly cached input at $0.075 per million — cached-input pricing down 70%, Alibaba says. The Work360 and token-use figures are Alibaba's own; the list prices match QwenCloud's model pages. Sanchit Vir Gogia, chief analyst at Greyhound Research, told InfoWorld that enterprises should treat the cost figure as a vendor benchmark rather than an assured reduction in their own AI spending.

What outside scores and the safety work show

Independent launch-day boards put Qwen4-Max ahead of its predecessor, with room still above it. Artificial Analysis scored Qwen4-Max at xhigh reasoning at 43 on its Intelligence Index, nine points ahead of Qwen3.8-Max and three points behind Kimi K3.1 Pro at 46, at about $3.27 per Index task. On Artificial Analysis's Terminal-Bench 4.0 result the model scored 48%, against 67% for Kimi K3.1 Pro — a different measure from Alibaba's token-use reduction on the same benchmark. In LM Arena's Code Arena WebDev board it debuted at No. 6 with a preliminary 1,750 points, 74 points above Qwen3.8-Max, and Arena scores are early and can move.

Safety was the other half of the launch, and Reuters led with it. Alibaba says it added code security checks alongside functionality checks in training: on Last Secure Code Bench, the share of functions passing both functionality and security tests rose from 22.0% for Qwen3.8-Max to 31.8%, and code secure at function, file and repository level at once rose from 8% to 14.7%. For cybersecurity training data, the company says it kept defensive material such as vulnerability verification and remediation and removed offensive material, and strengthened safeguards against chemical, biological, radiological and nuclear misuse; its internal testing found no evidence of significant high-risk capabilities. Reuters said it had not independently verified the findings, and analysts quoted by CSO Online advised treating both the cost and safety figures as claims to test, not results to assume.

Availability and limits

Qwen4-Max is available now, not just announced: the public API is live on QwenCloud under the model ID qwen4-max, after a preview that ran on Alibaba's Token Plan only from November 11, and the model is also in Qwen Studio, QwenWork and Qoder. Alibaba's showcases point at long, multi-step work — in one manufacturing demonstration with partners SANF and Pacific Union, manual programming time for a single workpiece fell from about 475 minutes to 24 minutes — but Alibaba describes these as showcases, and they have not been independently verified.

Qwen4-Max itself is not released as open weights: Alibaba says open weights for smaller Qwen4 models will be released, with no date given. The 62% figure is an average on Alibaba's own Work360 tasks; Gogia told InfoWorld that the relevant measure for a buyer is the cost of an accepted result, including retries, reasoning tokens and employee time spent checking outputs.

Disclaimer: Derrington Collaborative AI is my personal business, operated independently of and outside my official U.S. Government position and duties. The views and opinions expressed on this site and on any and all of my social media accounts are my own. They do not reflect the official policy or position of the Department of the Army, the Department of Defense, or the U.S. Government. Mention of any company, product, or service does not constitute an official endorsement.