Alibaba's Qwen4-Max Packs 2.4 Trillion Parameters and Cuts Task Costs by Over 60%
The Chinese AI labs aren't slowing down. Alibaba released Qwen4-Max on Monday, its new flagship model, and the pitch isn't only "smarter". It's "a lot cheaper to get real work done".
The model has 2.4 trillion parameters. CASRAI, citing Alibaba's own launch post, says it uses a mixture-of-experts design, meaning only about 71 billion of those parameters are active for any given word. It also handles images natively and can take in about 1 million tokens of context, enough to read a stack of long documents at once. A limited preview went out on November 11, and now it's generally available through Alibaba's QwenCloud API and its QwenWork and QwenStudio apps. Listed API pricing is $1.50 per million input tokens and $4.50 per million output tokens. Don't expect to download it, though. The Max model isn't being released as open weights, although smaller open Qwen4 models are promised later.
The headline number comes from Alibaba's own internal test, called Work360. Analytics India Magazine reports the average cost per task dropped from $4.43 with the previous Qwen3.8-Max to $1.67, about 62% less. It also used fewer tokens to get there.
Alibaba is also leaning into safety, which is a bit new for a Chinese lab's marketing. It says it kept training data that helps the model find and fix security holes but removed data that would make it better at attacking systems. It also added safety training around chemical, biological and nuclear risks. On one secure-coding benchmark, the share of code that both worked and was secure climbed from 22% to 31.8%. That's an improvement, but it also means most code still didn't pass.
Keep in mind these are Alibaba's own numbers, and the benchmarks shared didn't compare it directly to OpenAI, Google or Anthropic models.
