Skip to content
DealsBlack Friday & Cyber Monday · 19-30 November · tech deals, re-checked every hourSee the deals
Newsletter
News

Alibaba Launches Qwen4-Max AI Model With Tighter Safety Controls

Alibaba's new flagship AI model keeps the predecessor's 2.4 trillion parameters but activates fewer per token, costs 25% less on headline API rates and was trained with security and weapons-risk filtering.

Alibaba Launches Qwen4-Max AI Model With Tighter Safety Controls

What you need to know

  • Alibaba launched Qwen4-Max on 16 November 2026. It is available now through QwenCloud, QwenWork and QwenStudio.
  • It has 2.4 trillion total parameters, activates 71 billion per token, and has a one-million-token context window. It understands text, images and video natively.
  • International API pricing is $1.50 per million input tokens and $4.50 per million output tokens, 25% below Qwen3.8-Max. No UK price in pounds has been confirmed.

Alibaba has launched Qwen4-Max, which it calls the most capable model in its Qwen family so far. The company announced the model in Hangzhou on Monday 16 November and describes it as built for complex professional, creative, coding, scientific and agent-based work. It also says safety controls were built into both its training and its deployment.

Server racks lit in blue inside a data centre
Qwen4-Max arrives less than four months after Qwen3.8-Max, which Alibaba launched on 3 August 2026.

The model is available now through QwenCloud, via Token Plan and pay-as-you-go API access, as well as through QwenWork and QwenStudio. According to Reuters, Alibaba said:

"It can take on a wide range of creative and professional work"

Same size, leaner running

Qwen4-Max has the same 2.4 trillion total parameters and one-million-token context window as its predecessor, Qwen3.8-Max. It uses a sparse Mixture-of-Experts design, which means only part of the model is used for each token it processes. Alibaba says this part has shrunk from 95 billion activated parameters to 71 billion, a 25% reduction.

Other headline specifications include:

  • Native understanding of text, images and video
  • Support for deep reasoning, coding, tool calling, web search and structured outputs
  • Compatibility with the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages APIs
  • A maximum combined thinking-and-answer output of 262,144 tokens, according to QwenCloud's documentation for the production model

The production model, identified as qwen4-max, replaces a qwen4-max-preview version that entered preview on 11 November. The preview listed a higher output ceiling of 393,216 tokens. Alibaba Cloud's documentation now recommends the production version and schedules the preview for retirement at midnight on 25 November, China time (UTC+8).

Alibaba's performance claims

Alibaba says Qwen4-Max ranked sixth on Code Arena and improved on Qwen3.8-Max on Work360, the company's own benchmark covering 721 specialised fields. In Work360's high-effort reasoning mode, Alibaba says the new model used 35% fewer tokens and cut the average cost per task by about 60%. These figures come from Alibaba and have not been independently assessed.

The company also cited several demonstrations:

  • A cloud-resource-scheduling test cut data-annotation model-call costs by an average of 9% and made offline CPU rescheduling more than 14 times faster.
  • A manufacturing test reduced manual workpiece-programming time from about 475 minutes to 24.
  • A 48-hour scientific agent run produced an experimentally validated coolant molecule.
  • An Alzheimer's research experiment involved 454 tool calls, 3,239 metric evaluations and 24 candidate molecules.

Safety measures

Alibaba says its coding training checked for security as well as for working code. The aim was to reduce vulnerabilities such as code injection, memory-safety flaws and broken access controls. Cybersecurity training data was filtered by intent: material that helps detect and fix vulnerabilities was kept, while offensive attack content was removed. Additional safety datasets and post-training work targeted risks involving chemical, biological, radiological and nuclear materials.

Reuters reported Alibaba as saying internal testing found no evidence of significant high-risk capabilities. The company has not published the methodology or detailed results of that testing.

Pricing and availability

QwenCloud's price page lists international deployment of Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens. Cached-input reads cost $0.075 per million tokens. Cache writes cost $1.875 per million tokens for five-minute caching and $3 for one-hour caching.

Global deployment is slightly cheaper:

  • Input: $1.413 per million tokens
  • Output: $4.24 per million tokens
  • Cached-input reads: $0.071 per million tokens
  • Five-minute cache writes: $1.767 per million tokens
  • One-hour cache writes: $2.827 per million tokens

The headline input and output rates are 25% lower than Qwen3.8-Max's international pricing of $2 and $6. Prices are quoted in US dollars only. A sterling price, VAT treatment and any UK-specific subscription have not been confirmed.

QwenCloud lists deployment regions in Singapore, Frankfurt, Hong Kong, Tokyo and Virginia. A dedicated UK region has not been confirmed.

What's next

Alibaba says smaller open-weight Qwen4 models will be released to developers worldwide in future, though it has not given dates. Weights for Qwen4-Max itself have not been confirmed. That is a change from August, when Alibaba said Qwen3.8-Max's weights would follow the week after launch. Qwen4-Max arrives less than four months after that model, continuing the Qwen line's rapid release pace.

Why it matters

For UK developers and businesses, the main draw is a cheaper flagship model that works with the OpenAI and Anthropic API formats, so testing it against existing tools should need little rework. There are caveats. Prices are listed only in US dollars, VAT treatment is unconfirmed, and there is no confirmed UK hosting region, which matters for organisations with data-location requirements. The performance and safety claims come from Alibaba itself and have not been independently assessed. Unlike Qwen3.8-Max, there is no confirmed plan to release the flagship's weights for developers to run themselves.

Sources and evidence (10)
↑↓ to moveEnter to openEsc to close
GadgetScout

What do you need?

Browse by gadget

Topic hubs

Shop by category

More