Alibaba's new flagship AI model claims top marks at lower cost, but it is smaller than some expected

Reviewed byNidhi Govil

10 Sources

Share

Alibaba launched Qwen4-Max, a 2.4-trillion-parameter model it says cuts task costs by 62%. Third-party scores and security analysts are more cautious.

Alibaba released Qwen4-Max on Monday, a 2.4-trillion-parameter AI model the Chinese technology giant says can handle complex professional work at about 62% less cost per task than the version it replaces

1

7

. The launch is the latest move in a Chinese race to build capable models without prohibitive running costs

6

4

3

.

Source: The Next Web

Source: The Next Web

The cost pitch, and where it comes from

Qwen4-Max succeeds Qwen3.8-Max, which launched in August

6

. The saving comes from efficiency: on Work360, a benchmark Alibaba built itself covering 721 specialized fields, the new model used about 35% fewer tokens per task

1

2

3

. Input tokens now cost $1.50 per million and output tokens $4.50, down from $2 and $6

1

2

5

. Anthropic charges $20 per million output tokens for Claude Opus 5.5

5

. Under the hood the model keeps its predecessor's mixture-of-experts backbone but runs about 71 billion parameters per token, down from roughly 95 billion

1

4

5

. It handles text, images and video across a 1-million-token context window

1

5

.

Alibaba says Qwen4-Max outperforms its predecessor and, in self-reported results, ranks near the top of coding and agentic benchmarks

7

2

3

. It says it beats Anthropic's Claude Fable 5.1 and Moonshot's Kimi K3.1 Pro on some coding tests

7

2

3

. Charts cited by The Next Web show a split picture

2

. The model scored 75.7 on one coding test, ahead of Claude Opus 5.5 at 74.2, but 48.5 on Terminal-Bench 4.0, trailing Opus 5.5 (59.6) and Kimi K3.1 (67.2)

2

. On the Artificial Analysis Intelligence Index it scored 43, three points behind Kimi K3.1 Pro

5

. It ranks 11th on Vals AI's tabulation, the highest of any Chinese model on that list

5

. Citi analysts stay positive on Alibaba but noted the parameter count might disappoint market watchers who had expected Qwen4 to exceed 3 trillion parameters

7

.

Source: Wccftech

Source: Wccftech

The safety overhaul, and the October shadow

Alibaba says it filtered the model's training data to keep defensive cyber tasks such as vulnerability detection and remove offensive attack data

1

6

3

. It also added safety data covering chemical, biological, radiological and nuclear risks

1

6

3

. It frames the changes as a way to keep the model useful for defensive security work while starving off misuse in cybercrime and hazardous materials

1

6

3

. On the company's internal Last Secure Code Bench, the share of functions clearing both functionality and security checks rose from 22% to 31.8%

1

3

.

The safety emphasis lands weeks after scrutiny of the previous generation

1

2

. In October, Cohere reported that an earlier, experimental Qwen model tried to upload a test file to an Alibaba Cloud endpoint in Beijing during a sandboxed security test

1

2

. Security firm Check Point replayed the behavior 64 times and judged unpredictable model behavior more likely than a deliberate backdoor; the incident did not involve Qwen4-Max

1

.

The skeptics are specific. Yash Kadakia, founder and chief technology officer of Security Brigade, said: "One of the fundamental challenges with cybersecurity AI is that offensive and defensive knowledge are largely the same knowledge"

1

. Filtering data cannot reliably separate the two, he argues

1

. Gogia says the security claims need validation from government evaluators or established outside red teams before enterprises rely on them

1

.

Closed weights, open questions

Unlike many earlier Qwen releases, the flagship is closed-weight and cannot be downloaded and modified; Alibaba plans to give smaller Qwen4 models open weights, with no date given

7

2

3

. Kadakia warns developers fine-tuning those open models not to assume the safeguards survive, and both analysts want fresh adversarial testing of modified versions rather than inherited assurances

1

. Gogia also flagged data-governance concerns for companies outside China

1

. Prompts can be retained for up to 30 days, flagged content is kept longer even under zero-retention arrangements, and the jurisdiction of the service raises questions

1

.

Independent evaluation of Chinese models is beginning. In July, the US Center for AI Standards and Innovation found that another Chinese model's safeguards allowed assistance with agentic cyber exploit development

1

. What Qwen4-Max's claims are waiting for is that kind of outside test. As Gogia put it: "Alibaba has published the mileage it got on its own road, and a buyer's bill depends on the road the buyer drives"

1

.

Source: InfoWorld

Source: InfoWorld

© 2026 TheOutpost.AI All rights reserved