
Key Points
- 01Alibaba launches Qwen4-Max, a 2.4-trillion-parameter AI model
- 02Average task cost on Alibaba’s Work360 benchmark falls about 60%
- 03Token prices cut, with cached input as low as $0.075 per million
- 04New safety guardrails target cyber and CBRN risk mitigation
Alibaba launches Qwen4-Max flagship AI model
Alibaba Group has introduced Qwen4-Max, described as its most advanced artificial intelligence model and the first flagship in the Qwen4 series. The model’s release is framed around significantly lower usage costs, with Alibaba stating that average cost per task on its internal Work360 benchmark has fallen by about 60% compared with the previous Qwen3.8-Max generation. Qwen4-Max is positioned as a high-performance foundation model aimed at enterprise workloads, combining greater capability with reduced inference expense.
Qwen4-Max builds on a backbone of 2.4 trillion parameters, the same overall scale as Qwen3.8-Max, but it employs a sparse mixture-of-experts design that activates 71 billion parameters per token. This approach reduces the number of active parameters by about 25% versus the earlier Qwen3.8-Max-Base while seeking to maintain or improve reasoning performance. The model also offers native image and video understanding and supports a context window of 1 million tokens for long documents and complex workflows.
On the Work360 benchmark, which covers 26 primary domains and hundreds of professional subfields, official comparison data show the average cost per task dropping from 26.58 yuan to 11.16 yuan. Costs fell from $4.43 to $1.67 per task, a 62% reduction.
Lower token prices and higher efficiency
In addition to architectural changes, Alibaba has reduced list prices for using Qwen4-Max. Internationally, input tokens are priced at $1.50 per million and output tokens at $4.50 per million, both lower than the rates for Qwen3.8-Max. Cached input reads cost $0.075 per million tokens, while cache-hit pricing is 0.5 yuan (about $0.07) per million tokens.
Token efficiency gains contribute further to cost reductions. On the Work360 benchmark, Alibaba reports that Qwen4-Max uses about 24% fewer tokens at its highest reasoning intensity and about 35% fewer tokens at a high reasoning level, compared with Qwen3.8-Max. On the Terminal-Bench 4.0 evaluation, token consumption is reduced by roughly 48% at Xhigh and roughly 49% at High reasoning settings.
Performance scores on internal metrics also improve alongside cost. On Work360, Qwen4-Max scored 52.69 compared with 47.74 for Qwen3.8-Max. Qwen4-Max ranked sixth in the CodeArena blind test for front-end coding and placed in the leading tier on agentic and professional task evaluations such as Agents’ Last Exam and the Vals Index, supporting its positioning for software development and long-horizon professional use cases.
Cybersecurity and CBRN safety guardrails
Alibaba highlights tighter safety measures as a core feature of Qwen4-Max. For cybersecurity, the company says it filtered training data based on task purpose, retaining data for defensive activities like vulnerability verification, detection and remediation, while removing content that would enhance offensive cyberattack capabilities. An internal expert review of long agent runs found no capabilities that pose large-scale cybersecurity risks.
On secure coding tests, Qwen4-Max is said to increase the share of functions passing both functionality and security checks to 31.8%, up from 22.0% for Qwen3.8-Max. Beyond cyber, Alibaba reports adding dedicated safety data during post-training to address chemical, biological, radiological and nuclear risks.
Access, products and open-source plans
Qwen4-Max is being made available through multiple Alibaba channels. Developers can access the model via API on the Qianwen AI Platform and through QwenCloud, including via a Token Plan subscription. A limited preview of the model became available through the Token Plan on November 11, and it is also integrated into applications such as QwenWork, Qoder, Qwen Studio, the Qianwen app and Qianwen Office for end users.
While Qwen4-Max itself is a closed-weight flagship, Alibaba reiterates its commitment to an open-source strategy. The company states that smaller and mid-sized models in the Qwen4 series will be released as open source in stages, with model weights to be made public.
Key Takeaways
- 01Qwen4-Max marks a step change in Alibaba’s AI line-up by pairing a 2.4-trillion-parameter backbone with a sparser activation scheme to cut active compute per token.
- 02Across Alibaba’s Work360 benchmark, Qwen4-Max reduces both token usage and list prices, resulting in an average per-task cost that is roughly 60% lower than Qwen3.8-Max.
- 03Enhanced cybersecurity and CBRN guardrails, along with stronger secure coding results, show Alibaba positioning Qwen4-Max as a safer high-end model while keeping smaller Qwen4 variants on an open-source track.
References
- https://stocktwits.com/news-articles/markets/equity/baba-stock-edges-lower-overnight-alibaba-reportedly-launches-most-advanced-ai-model-that-cuts-task-costs-by-60/cNb7gKGR9Xf?.tsrc=rss
- https://www.scmp.com/tech/big-tech/article/3374540/alibaba-says-new-flagship-ai-model-cuts-task-costs-60-while-tightening-cyber-guardrails
- https://mlq.ai/news/alibaba-launches-qwen4-max-saying-it-cuts-task-costs-by-about-60-with-tighter-cyber-guardrails/
- https://finance.biggo.com/news/357f82fc-873e-4025-8c0e-f2b920922b2b