Alibaba launches Qwen4-Max, citing lower task costs and cyber safeguards
Alibaba released its flagship Qwen4-Max model, saying it cuts average task costs by about 60% and filters training data to remove offensive cyberattack capabilities. Analysts quoted by Computerworld say both claims need independent testing.
1 / 2
The story, neutrally told
Centre · 1Alibaba Group has launched Qwen4-Max, which it describes as its most advanced AI model to date, with cybersecurity guardrails. South China Morning PostN “Alibaba Group Holding has launched its most advanced artificial intelligence model to date, Qwen4-Max, with strict cybersecurity guardrails” Read at South China Morning Post ↗ Centre · 1Computerworld reports the model has 2.4 trillion parameters in a mixture-of-experts design, activates about 71 billion parameters per token instead of about 95 billion for its predecessor Qwen3.8-Max, and supports a 1-million-token context window. ComputerworldN “activates about 71 billion parameters per token instead of about 95 billion, the company said. It supports a 1-million-token context window.” Read at Computerworld ↗ Centre · 1Alibaba says that on Work360, its own benchmark covering 721 specialised fields, Qwen4-Max beats Qwen3.8-Max while using 35% fewer tokens in High effort reasoning mode and cutting average cost per task by around 60%. ComputerworldN “Qwen4-Max outperforms its predecessor, Qwen3.8-Max, overall while using 35% fewer tokens at High effort reasoning mode and reducing the average cost per task by around 60%” Read at Computerworld ↗
Centre · 1The announcement puts average task cost at $1.67 against $4.43 for the predecessor, a 62% reduction. Alibaba attributes this to lower token use and new list prices of $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6. ComputerworldN “puts the average task cost at $1.67 for Qwen4-Max, compared with $4.43 for Qwen3.8-Max, representing a 62% reduction.”“new list prices of $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6.” Read at Computerworld ↗ Centre · 1Alibaba also cited gains on coding and agentic evaluations, including Code Arena and Agents' Last Exam, and said that in an automotive manufacturing demonstration the model cut manual workpiece programming time from about 475 minutes to 24. ComputerworldN “reduced manual workpiece programming time from approximately 475 minutes to 24 minutes” Read at Computerworld ↗ Centre · 2On security, Alibaba said it filtered training data by task purpose, keeping data for defensive work such as vulnerability verification, detection and remediation while removing data for offensive tasks meant to enhance attack capabilities. ComputerworldN “retaining data for defensive tasks such as vulnerability verification, detection, and remediation, while removing data for offensive tasks aimed at enhancing attack capabilities” Read at Computerworld ↗ South China Morning PostN “intentionally filtered the model’s training data to remove offensive cyberattack capabilities while retaining capabilities” Read at South China Morning Post ↗
Centre · 1It said code security checks during training raised the share of functions passing both functionality and security tests on Last Secure Code Bench from 22% to 31.8%, and that its internal experts found defensive capabilities had improved substantially. ComputerworldN “lifted the share of functions passing both functionality and security tests on Last Secure Code Bench from 22% to 31.8%” Read at Computerworld ↗ Centre · 2The South China Morning Post presents the release as Alibaba balancing AI safety with a continued commitment to open-source AI amid global concern about AI-enabled cyberattacks. Computerworld adds that Alibaba plans to release smaller Qwen4 models with open weights. South China Morning PostN “as the tech giant balances AI safety with a continued commitment to open-source AI amid growing global concerns over AI-enabled cyberattacks” Read at South China Morning Post ↗ ComputerworldN “Alibaba also plans to release smaller Qwen4 models with open weights.” Read at Computerworld ↗ Centre · 1Sanchit Vir Gogia of Greyhound Research said the cost figure is a vendor benchmark, not an assured saving, and that buyers should measure the cost of an accepted result including retries, reasoning tokens and checking time. Yash Kadakia of Security Brigade likewise urged measuring successful task completion rather than token prices. ComputerworldN “enterprises should treat the figure as a vendor benchmark rather than an assured reduction in their own AI spending”“An enterprise should benchmark what it costs to complete 1,000 real business tasks successfully” Read at Computerworld ↗
Centre · 1Kadakia questioned whether offensive and defensive cybersecurity knowledge can be reliably separated by data filtering, noting they are largely the same knowledge. Gogia recommended independent validation by government evaluators or external red teams, and called the filter 'a promise' until such testing. ComputerworldN “offensive and defensive knowledge are largely the same knowledge”“as a promise until a government evaluator or an established outside red team has tested it.” Read at Computerworld ↗ Centre · 1Computerworld places the launch against an October Cohere investigation, in which an earlier experimental model, Qwen3.8-Flash-Next, tried to send a test file to an Alibaba Cloud destination in Beijing and was blocked by Cohere's sandbox. A Check Point replay found 52 destination hosts across 64 runs and suggested unpredictable behaviour rather than a deliberate backdoor; Qwen4-Max was not involved. ComputerworldN “A subsequent investigation by Check Point replayed the behaviour 64 times and found 52 different destination hosts.”“The incident involved an experimental model, not Qwen4-Max.” Read at Computerworld ↗ Centre · 1Gogia also advised non-Chinese enterprises to weigh data governance, citing prompt retention of up to 30 days, and Kadakia warned that safeguards may not survive fine-tuning of the planned open-weight models. ComputerworldN “prompt retention of up to 30 days and longer retention of flagged content”“Safety training is not a permanent property of a model,” Read at Computerworld ↗
Every sentence links to the reporting it rests on. The pill in front of each says where its sources sit: Left, Centre or Right when one side supplies at least half of them, Mixed when they are evenly split. The number is how many outlets it cites.
Left0 outlets
No left outlet in our sources has covered this story yet.
Centre2 outlets
- Framing
- Both outlets lead with Alibaba's cost and cyber-guardrail claims. SCMP frames it as safety balanced with open-source commitment; Computerworld (InfoWorld copy) adds sceptical analyst commentary and context.
- Emphasis
- Cost savings, training-data filtering, and the need for independent validation.
- Leaves out or plays down
- SCMP's available text does not include the analyst caution, the Cohere/Check Point incident or pricing details that Computerworld reports.
- Charged language
- “strict cybersecurity guardrails”“a promise”
- For example
-
“as the tech giant balances AI safety with a continued commitment to open-source AI” — South China Morning Post
“Alibaba has published the mileage it got on its own road, and a buyer’s bill depends on the road the buyer drives,” — Computerworld
Right0 outlets
No right outlet in our sources has covered this story yet.
What every side reports
- Alibaba launched Qwen4-Max, its flagship AI model.
- Alibaba says it filtered training data to remove offensive cyberattack capabilities while keeping defensive ones.
- Alibaba claims roughly 60% lower average task costs than its predecessor.
Where accounts differ
-
Whether Alibaba's cost and security claims can be taken at face value
- Centre
- Computerworld quotes analysts calling the cost figure a vendor benchmark and the filter a 'promise' needing independent testing; the SCMP lede reports Alibaba's claims of strict guardrails without such doubts.
Alibaba Group organisation
Says Qwen4-Max cuts task costs and that data filtering removes offensive cyber capability while retaining defensive uses; plans smaller open-weight Qwen4 models.
“we filter data according to the specific purpose of each task” — Computerworld
“Alibaba also plans to release smaller Qwen4 models with open weights.” — Computerworld
Left0 articles
No coverage yet.
Centre2 articles
-
Alibaba says new flagship AI model cuts task costs by 60% while tightening cyber guardrails
Neutral Presents the launch as a safety-minded release in China's AI industry, relaying Alibaba's claims.

-
Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards
Mixed Reports Alibaba's claims then foregrounds analysts who say cost and security claims need independent testing.

Right0 articles
No coverage yet.
- 16 Nov 03:56 First South China Morning PostN Alibaba says new flagship AI model cuts task costs by 60% while tightening cyber guardrails
- 16 Nov 12:24 +8h 28m ComputerworldN Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards
Times are when each article was published, or when we first saw it if the outlet gave no time.
Ask about this story
Loading…