
Alibaba's Qwen4-MAX cuts task costs 62% to $1.67, company says
Alibaba says Qwen4-MAX cuts task cost 62% to $1.67 on its own Work360 benchmark, activates 71B of 2.4T parameters per token, and ships without published weights.
- By
- Sophie Lindqvist
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
Alibaba released Qwen4-MAX on Nov. 16 and claims average cost per task fell 62% against its predecessor, from $4.43 to $1.67, on the company's own Work360 benchmark. The model carries a 2.4-trillion-parameter backbone, native image and video understanding, and a one-million-token context window. It is the first release in the Qwen4 family. The Qwen team positioned it plainly: the new model does more work for less money than Qwen3.8-MAX.
Every headline figure here comes from Alibaba itself. The company has not published independent verification, and Work360 is an Alibaba-built benchmark. That caveat belongs next to every number below, not in a footnote.
Where do the savings come from?
Two levers drive the claimed reduction: fewer tokens consumed per task, and a lower price per token. Alibaba says Qwen4-MAX used about 24% fewer tokens than Qwen3.8-MAX on Work360 at its highest (Xhigh) reasoning effort, and 35% fewer at High effort. On Terminal-Bench 4.0, the reduction reached about 48% to 49%.
List prices dropped as well. International prices from Singapore, via Alibaba Cloud Model Studio:
- Input: $1.50 per million tokens, down from $2.00
- Output: $4.50 per million tokens, down from $6.00
- Cached input reads: $0.075 per million tokens for Qwen4-MAX
The South China Morning Post reported the cost cut at about 60%. On Work360, which covers 26 top-level work domains and 721 specialized fields, Qwen4-MAX scored 52.69, up from 47.74 for Qwen3.8-Max. Again, these scores come from Alibaba's own evaluation suite.
What changed under the hood?
Qwen4-MAX is a sparse mixture-of-experts design. Alibaba says it activates 71 billion parameters per token out of the 2.4-trillion backbone, plus 328 billion N-gram embedding parameters. The architecture mixes Gated DeltaNet layers with a new sparse attention scheme intended to cut the cost of long contexts. The pre-trained base model activates 25% fewer parameters than Qwen3.8-MAX-Base, according to the company.
How did Alibaba handle cyber and CBRN risks?
The company says it filtered training data by the purpose of each cybersecurity task. It kept data for defensive work, such as vulnerability verification, detection and remediation. It removed data meant to enhance attack capabilities. An internal expert review of long agent runs found no capabilities posing large-scale cybersecurity risks, Alibaba said. Internal review is the operative phrase; no third-party assessment accompanies the launch.
For chemical, biological, radiological and nuclear risks, Alibaba says it added dedicated safety data in post-training. Internal experts concluded the model's answers did not give users a significant uplift, according to the company.
One security metric did improve on Alibaba's own testing. On a secure-coding test, the share of functions that passed both functionality and security checks rose from 22.0% for Qwen3.8-Max to 31.8%. A passing grade for fewer than a third of functions leaves clear room for the next iteration.
Closed flagship, open small models later
Access is API-only. Qwen4-MAX is available through QwenCloud's API and Token Plan subscription, and inside Alibaba's QwenWork, Qoder and Qwen Studio products. Alibaba did not release the model's weights. A limited preview opened through the Token Plan on Nov. 11, five days before the general release.
The company says smaller Qwen4 models will arrive as open source. It gave no date. A roadmap without a schedule is a promise, not a plan, and the open-source community will judge the Qwen4 family by what ships next.
For now, Alibaba's claim rests on a self-built benchmark, self-reported prices and self-conducted safety reviews. If independent evaluations confirm even half the claimed efficiency gain, the $1.50-per-million-token input price will press directly on rival flagship APIs.
Source: mlq.ai
More from Sophie Lindqvist
Show full bio
News editor covering money, funding and market shifts around Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,055 articles
Related articles
digi-international-s-q4-25-revenue-growth-meets-a-softer-road-ahead
Digi International's Q4: 25% Revenue Growth Meets a Softer Road Ahead
ibm-claims-quantum-advantage-with-three-heron-chip-demonstrations-52c38ac2
IBM Claims Quantum Advantage With Three Heron Chip Demonstrations
cerebras-targets-cs-5-production-as-backlog-hits-26-1-billion
Cerebras Targets CS-5 Production for 2027 as Backlog Hits $26.1 Billion
microsoft-s-new-maia-ai-chips-promise-better-token-efficiency-61ef4494
Microsoft's New Maia AI Chips Promise Better Token Efficiency


