AI Card Bao AICARDBAO.COM

Industry News

Alibaba releases Qwen4-Max: 2.4T parameters, native vision, millions of contexts, cache hit 0.5 yuan per million tokens

Alibaba today launched Tongyi Qianwen’s new-generation flagship Qwen4-Max. The book parameters are very strong: 2.4T total parameters, native visual capabilities, and context support up to 1 million tokens.

The API has been opened on the Qianwen platform. The price is the most memorable part this time: the cache hit is as low as 0.5 yuan per million tokens. Ali’s own statement is that the average cost of running a task has dropped by 60% compared with the previous generation.

However, third-party figures are not that exaggerated. Artificial Analysis gave Qwen4-Max an intelligence index of 43 points under xhigh inference intensity, 9 points higher than Qwen3.8-Max, and the single-task cost is about 11% lower.

The gap between 60% and more than 10% is not small. The reason most likely lies in the measurement caliber: the official calculation is based on its own typical tasks and full cache discounts; the third party calculates unified question sets and unified reasoning levels. How much you can actually save depends on how many of your requests can hit the cache.

In terms of capabilities, it is 3 points lower than Kimi K3.1 Pro in the same list, and it is not yet ranked first in China. Another list is even more impressive: Image-to-WebDev Arena ranked sixth with 1674 points, 35 points more than the previous generation, ahead of Claude Opus 5, GPT-6 Sol and Muse Spark 1.3. This list is about reading pictures and writing web pages, which is where native visuals come in.

The official positioning is office productivity, creativity, software development and scientific research, focusing on tool invocation, long-term tasks and final delivery quality. The named scenarios include 3D modeling, game development, finance, law and medical care.

In terms of entrance, the preview version is first available on Alibaba Token Plan, with a minimum discount of 10% off for a limited time; Qoder, Qianwen App, and QwenStudio have been simultaneously accessed.

There is another sentence that is easily overlooked: small and medium-sized models of the Qwen4 series will be open sourced one after another. For teams deploying their own models, this sentence may be more useful than Max's running score.

Related