Architecture and size of Qwen4-Max
Alibaba has released Qwen4-Max as a proprietary model. According to the Qwen team, it has 2.4 trillion parameters but activates only 71 billion per token. It processes images natively and supports a context of one million tokens. These figures are reported by both The Decoder and MarkTechPost.
The Decoder describes the hybrid architecture of the Qwen4 series in more detail. Instead of exclusively standard attention layers, three layers compress the previous context into a condensed state and a fourth selects the parts that matter. This reduces computational cost on long inputs. Also added are four parallel connections between layers, with dynamic gates to stabilize the flow of information, and an N-gram embedding layer with 328 billion parameters.
On training, Alibaba says the 2.4-trillion-parameter configuration ran stably and the loss fell steadily, without sudden spikes. In internal measurements in bits per byte (knowledge, code, question answering, STEM), Qwen4-Max-Base had the best values among the five checkpoints tested. It improved by 5 to 13 percent over Qwen3.8-Max-Base while activating 25 percent fewer parameters per token.
Results: internal and independent figures
The Work360 figures come from Alibaba itself. Work360 is an internal benchmark of industrial tasks across 721 specialized domains. Qwen4-Max uses roughly a quarter to a third fewer tokens than its predecessor for the same tasks. On Terminal-Bench 4.0, an agent-task benchmark, token usage falls by almost half. The average cost per Work360 task drops from $4.43 to $1.67, or 62%, and the score rises from 47.74 to 52.69.
On external benchmarks the picture is mixed, as relayed by The Decoder from Alibaba's material. On Agents' Last Exam v1.1, for long-running agent tasks, the model reaches 46.7 points, the best result in the comparison group. It trails Opus 5.5 and GPT on Humanity's Last Exam. On HealthBench Professional it is behind the GPT models and ahead of Opus 5.5.
The independent measurement is from Artificial Analysis. At the highest reasoning setting the model scores 43 points on the Intelligence Index, nine more than Qwen3.8-Max and three fewer than Kimi K3.1 Pro, the top Chinese model with 46. The biggest gains over its predecessor appear on HLE-Diamond (an increase of 21 points, to 45 percent), AA-Omniscience (18) and AutomationBench-AA (17).
The index's cost per task is $3.27 for Qwen4-Max, against $10.02 for Qwen3.8-Max and $3.68 for Kimi K3.1 Pro. Tencent's Hy4 costs $1.03 per task, but scores eight points lower on the index. The savings come from three factors: token prices 25 percent lower, cache hits 70 percent cheaper and fewer output tokens, about 105,000 per task against 171,000 for Qwen3.8-Max.
Cost per task on the Artificial Analysis Intelligence Index and results on two code and agent benchmarks
| Model | Index cost per task | Terminal-Bench 4.0 | FrontierCode 2.0 |
|---|---|---|---|
| Qwen4-Max | $3.27 | 48% | 26% |
| Kimi K3.1 Pro | $3.68 | 67% | 26% |
| Tencent Hy4 | $1.03 | — | 31% |
Pricing and application areas claimed by Alibaba
Through Alibaba's API, input is charged at $1.50 per million tokens and output at $4.50. Cache hits cost $0.075. MarkTechPost also reports the input price.
The company presents a "Live Delivery" feature: the model delivers a usable first version early and extends it over later passes, while the user can step in at any moment. The related Qwen-Live-Delivery-Plugin is released as open source. Alibaba also says its automotive partners cut the manual programming time for a part from about 475 to 24 minutes. According to the company, in collaboration with the MindOpt solver it reduced the cost of allocating computing resources for data annotation tasks by an average of 10.42%.
In laboratory experiments, Alibaba says that with Deep Principle's Mira system the model operated autonomously for 48 hours and examined 2.7 million candidate molecules for data center processor coolant. With DP Technology's PharmMaster platform it optimized 24 candidate molecules for a compound related to an inflammatory response linked to Alzheimer's disease. Alibaba says the final candidate binds to the target protein about 3.6 times better than the starting compound. All of these figures come from Alibaba itself and have not been independently verified in the sources.