THE CRUNCH

Alibaba has launched Qwen4-Max, a flagship model it positions against Anthropic's Opus 5.5, OpenAI's GPT-6 and Moonshot's Kimi K3.1 Pro, with much lower running costs as the main selling point. The mixture-of-experts model has 2.4 trillion parameters but activates only 71 billion per token, according to the Qwen team, and processes images natively with a one-million-token context window.

The revamped Qwen4 architecture compresses prior context into a condensed state across three layers, with a fourth layer picking out only the relevant parts, which cuts compute on long inputs. An N-gram embedding layer with 328 billion parameters supplies common token sequences as ready-made building blocks. Alibaba says training of the full configuration ran stably, with loss falling steadily and no sudden spikes.

The cost story is where Alibaba's numbers are strongest. On its own Work360 benchmark for industrial tasks, the company says average cost per task fell 62 percent, from $4.43 to $1.67, while the score rose from 47.74 to 52.69. Artificial Analysis independently puts Qwen4-Max at $3.27 per Intelligence Index task, about a third of Qwen3.8-Max's $10.02 and roughly 11 percent less than Kimi K3.1 Pro's $3.68. Through its API, Alibaba charges $1.50 per million input tokens and $4.50 per million output tokens.

External results are more mixed. Qwen4-Max scores 43 on the Artificial Analysis Intelligence Index, three behind Kimi K3.1 Pro, and takes the best result in its comparison group on Agents' Last Exam v1.1 at 46.7 points. But it trails Opus 5.5 and the GPT models on Humanity's Last Exam, and on Terminal-Bench 4.0 it scores 48 percent against Kimi's 67 percent. On Cognition's FrontierCode 2.0 coding benchmark it ties Kimi at 26 percent and trails Tencent's Hy4 at 31 percent.

For agentic work, where users may only see results at the end of a long task, Alibaba introduces "Live Delivery": the model produces a usable first version early and refines it in later passes, letting users change course along the way. The accompanying Qwen-Live-Delivery-Plugin has been open-sourced. Alibaba also cites partner deployments, including automotive firms reportedly cutting manual programming time for a workpiece from about 475 minutes to 24, and internal use that reportedly cut compute-allocation costs for data annotation by an average of 10.42 percent.