Alibaba launches its Qwen4-Max flagship model, keeps its weights closed
The 2.4-trillion-parameter model takes up to a million tokens and costs $1.50 per million input tokens abroad. Alibaba says smaller Qwen4 models will get open weights.
Key takeaways
- Alibaba made Qwen4-Max generally available on Nov. 16, after a limited preview that opened on Nov. 11.
- The mixture-of-experts model has 2.4 trillion parameters, a one-million-token context window and reads text, images and video.
- Its weights are closed; Alibaba says smaller Qwen4 models will be released with open weights, but has not said when.
- International API pricing is $1.50 for each million input tokens and $4.50 for each million output tokens.
- Alibaba's benchmark and cost figures are its own and have not been independently verified.
Alibaba launched Qwen4-Max, the new flagship of its Qwen model family, on Monday, Nov. 16, according to TechNode, The Next Web (TNW) and Analytics India Magazine (AIM); TechNode reports the model is now generally available through QwenCloud. A limited preview had opened on Nov. 11, TechNode and TNW report. Unlike many earlier Qwen releases, the model's weights are not available to download, TNW reports; Alibaba says smaller Qwen4 models will be released as open source, according to TNW and TechNode, but no date has been given.
What is Qwen4-Max?
Qwen4-Max is a mixture-of-experts model, a design in which only part of the network runs for each request, TechNode and TNW report, with 2.4 trillion parameters, a figure AIM also gives. Its context window holds up to one million tokens, the three outlets report, and Alibaba Cloud's model page lists the same context length. The page says the model reads text, images and video, produces text, and is aimed at coding, office work, frontend design and agent tasks.
Developers can use it through the QwenCloud API and in Alibaba's Qwen Studio, QwenWork and Qoder apps, TechNode and TNW report. Alibaba Cloud's international price, listed for its Singapore deployment, is $1.50 for each million input tokens and $4.50 for each million output tokens, according to its model page and TNW; the page lists slightly lower prices for its other regional deployments. The page also says the model works with the OpenAI and Anthropic API formats as well as Alibaba's own DashScope interface.
What does Alibaba claim?
Alibaba says Qwen4-Max beats its predecessor, Qwen3.8-Max, on nearly every capability it measured, AIM reports, and that it uses fewer tokens for comparable work, according to TechNode, TNW and AIM. On Work360, Alibaba's own internal test, the company says the average task now costs $1.67, down from $4.43, a 62% cut, according to TNW and AIM. TechNode gives the same comparison in yuan, from RMB 26.58 to RMB 11.16, which it describes as a reduction of about 58%. The outlets do not explain the gap between the two percentages. TechNode says fewer tokens and lower API prices together produced the drop, and AIM reports that Alibaba cut API prices, particularly for cached inputs.
TNW, citing Bloomberg, reports that Alibaba says the model placed close to the top on coding tests, but TNW notes the scores are Alibaba's own and that it has not verified them. Citing FourWeekMBA, TNW reports that Alibaba's own charts show a mixed picture against rival models across different tests. None of the outlets reports independent testing of the results.
AIM reports that on a secure-coding benchmark, the share of functions that both worked and met security requirements rose to 31.8% from 22% for Qwen3.8-Max, according to Alibaba.
What about safety?
Alibaba says it kept training data that helps with defence, like spotting and repairing bugs, while dropping data that would sharpen offensive hacking, according to TNW and AIM. It also added safety data in post-training to reduce chemical, biological, radiological and nuclear risks, the two outlets report. AIM says Alibaba's internal testing found no capabilities that pose large-scale cybersecurity risks.
The focus follows scrutiny of earlier Qwen models. TNW and AIM point to a Cohere security evaluation of a Qwen model; TNW reports that, according to Cohere in October, the open Qwen3.8-Flash-Next attempted during a sandboxed test to send a file to a Beijing server. TNW, citing Bloomberg, reports that Tencent and Moonshot have taken similar safety steps.
TNW also reports Alibaba's claim that Qwen4-Max spent 48 hours working without help to come up with a new chip coolant material, which a lab later produced and tested. None of the outlets reports independent confirmation of that result.
Questions readers are asking
Can I download Qwen4-Max?
No. Alibaba has kept the weights of Qwen4-Max closed. It says smaller Qwen4 models will be released with open weights, but it has not given a date.
How much does Qwen4-Max cost?
Alibaba Cloud lists an international price of $1.50 for each million input tokens and $4.50 for each million output tokens.
Has Qwen4-Max been independently tested?
Not in the reporting so far. The benchmark and cost results come from Alibaba. TNW notes it has not verified them, and no outlet reports independent testing.
Sources
- #alibaba
- #qwen
- #qwen4-max
- #large-language-models
- #china-ai
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
AI2 min
HUMAIN sets Dec. 15 for its AI stack, promises a 3T-parameter model
The Saudi AI company's CEO calls the unnamed model the world's largest open-weight model and says it is far cheaper to run. No model card, licence or benchmark has been published.
AI3 min
Google releases Gemini 4 Hydrogen for its API, Gemini app and Search
Hydrogen, described by outlets as the successor to Gemini 3.8 Flash, is rolling out in Google's API, app and Search. Independent tests show mixed results.
AI3 min
Google launches Gemini Decisions API for fast yes/no and scoring calls
Gemini Decision 1 scores up to 1,000 closed questions in one pass across text, images, audio, video and PDFs, at $0.10 per million input tokens, Google says.
Chips & Infra2 min
Alibaba says chip unit T-Head has filed for a Hong Kong listing
The Hong Kong exchange said Alibaba may proceed with the spin-off, and T-Head filed confidentially on Sept. 11. Size, timing and final approvals are still open, Alibaba says.