A task costs sixty percent less on average with Qwen4-Max, Alibaba says, as Reuters points to Chinese training on OpenAI answers
Alibaba's new top model has been available through its own cloud since Monday, but its weights remain closed for now.
Alibaba on Monday made its new flagship model Qwen4-Max generally available. According to CASRAI, a limited group of customers on the so-called Token Plan had access to a preview version since 11 November. The model can now be used through the QwenCloud API and in the QwenWork and QwenStudio apps, and according to Jawlah it can also process texts of around one million tokens in one go. The price is 1.50 dollars per million input tokens and 4.50 dollars per million output tokens, with a much lower rate for text that is already cached.
SiliconANGLE describes how the model is built. It is a mixture-of-experts with 2.4 trillion parameters, of which 71 billion are active at any time, supplemented by 328 billion parameters that serve as a kind of memory for frequently repeated calculations. Two changes to the attention mechanism are meant to make long prompts cheaper. On an internal test for office work, Qwen4-Max used up to 35 percent fewer tokens than its predecessor Qwen3.8-Max, and on the Terminal-Bench 4.0 coding test as much as 49 percent fewer.
Jawlah lists the company's figures: on that internal Work360 test, the average cost per task fell from 4.43 to 1.67 dollars and the score rose from 47.74 to 52.69 points. On the Code Arena programming leaderboard the model ranks sixth. Both sites warn that these are all Alibaba's own measurements. The weights of Qwen4-Max cannot be downloaded; smaller open models on the same basis will follow later, but there is no date.
A day later, Reuters published an investigation into more than 300 Chinese documents, from academic papers to patents and job postings. According to the news agency, they show that Chinese developers use answers from American models such as GPT-4o on a large scale to further train their own models, including an older Qwen version in a 2024 Shanghai study. A US official calls this so-called distillation the main reason China is closing the gap, while sources at Chinese developers dispute that. Alibaba did not respond to Reuters' questions.