Alibaba released Qwen 4 Max on Monday, presenting the company as the leader in China’s increasingly competitive market for AI agents that can complete professional work over extended sessions. The model is the first flagship built on the new Qwen 4 architecture, combines text, image and video inputs, and accepts up to one million tokens of context. Its central pitch combines higher benchmark scores with sharply lower operating costs, a combination that could matter more to enterprise adopters than another rise in general model rankings.
Qwen 4 Max: A large model with a narrower active footprint
According to The Decoder, Qwen 4 Max uses a mixture-of-experts design with 2.4 trillion parameters while activating only 71 billion per token. Most of the network therefore remains idle while the model processes each word. SiliconANGLE reported the same 71 billion active parameters and described the model as Alibaba’s newest and most capable large language model. The effect is lower computing power per token without sacrificing the scale of the overall system.
That scale matters for tasks that involve lengthy codebases or large stacks of documents, Decrypt reported. The one-million-token context window gives the model considerable room to hold such material. Its visual capability is aimed at understanding images and video and then producing code or other text-based output, rather than serving as a new image generator, Decrypt said. Alibaba has positioned the model for professional work in industry, science and finance, according to The Decoder.
Agent results are promising but not uniformly dominant
On Agent Arena, Qwen 4 Max placed 14th among individual models and moved Alibaba to fifth place among laboratories, Decrypt reported. The leaderboard assesses real agent sessions through task completion, user satisfaction and the ability to respond to corrections. Qwen 4 Max posted a preliminary score of 7.46 percent across 14,727 sessions, the highest among Chinese entries. American laboratories still hold the top positions in Arena’s overall agent and web development tables, Decrypt noted.
The web development results showed a sharper improvement. On the Code Arena ranking, Qwen 4 Max debuted sixth globally with 1,750 points, an increase of 74 from its predecessor, Decrypt reported. It placed eighth in the branding and reference-based design categories. The ranking compares websites built by AI models through head-to-head user votes. Decrypt also cautioned that Alibaba’s demonstrations, including game environments and the conversion of factory layouts into 3D models, were company-selected examples rather than independent tests.
External benchmark results remain mixed
At its highest reasoning setting, Qwen 4 Max scores 43 points on the current Artificial Analysis Intelligence Index, The Decoder reported. That is nine points above Qwen 3.8 Max and three behind Kimi K3.1 Pro, the leading Chinese model at 46 points. The largest gains over the predecessor came on HLE-Diamond, where the score rose 21 points to 45 percent, along with gains of 18 on AA-Omniscience and 17 on AutomationBench-AA.
Results on longer agent tasks produced a strong claim. On Agents’ Last Exam v1.1, Qwen 4 Max scored 46.7 points, the best result in the comparison group, The Decoder reported. The same source said it trails Opus 5.5 and the GPT models on Humanity’s Last Exam, while on HealthBench Professional it falls behind the GPT models but edges past Opus 5.5. In coding tests, Qwen 4 Max scored 48 percent on Terminal-Bench 4.0, behind Kimi K3.1 Pro at 67 percent. On Cognition’s FrontierCode 2.0, it tied Kimi at 26 percent and trailed Tencent Hy4 at 31 percent.
Cost efficiency is the stronger story
Artificial Analysis puts the cost of Qwen 4 Max at 3.27 dollars per Intelligence Index task, The Decoder reported. That is about one third of the 10.02 dollars charged for Qwen 3.8 Max and roughly 11 percent below the 3.68 dollars for Kimi K3.1 Pro. Qwen 4 Max needs about 105,000 output tokens per benchmark task, compared with 171,000 for Qwen 3.8 Max. Tencent’s Hy4 is cheaper still at 1.03 dollars per task, but scores eight points lower on the index.
Alibaba’s own prices also moved lower. Qwen 4 Max costs 1.50 dollars per million input tokens and 4.50 dollars per million generated tokens through QwenCloud, Decrypt reported. That represents a 25 percent reduction from the previous flagship’s 2 dollars and 6 dollars rates. Reused input served from an automatic cache costs 0.075 dollars per million tokens. The Decoder described those three changes, lower token prices, cheaper cache hits and reduced token consumption, as the sources of the broader efficiency gains.
Internal work benchmarks favor the new model
Alibaba compared Qwen 4 Max with Qwen 3.8 Max using an internal benchmark that includes knowledge work tasks. SiliconANGLE reported that the new model achieved a higher score and used up to 35 percent fewer tokens, with savings growing to 49 percent on Terminal-Bench 4.0. Alibaba’s own Work360 benchmark covers industrial work tasks across 721 specialized fields, The Decoder said. On that test, average cost per task fell 62 percent, from 4.43 dollars to 1.67 dollars, while the score rose from 47.74 to 52.69.
Decrypt summarized Alibaba’s internal testing across hundreds of professional specialties, which found that Qwen 4 Max used 35 percent fewer tokens and reduced average task costs by roughly 60 percent compared with Qwen 3.8 Max. These are company claims based on internal measurements rather than independent verification. Even so, the direction of the reported findings aligns with the external cost data published by Artificial Analysis.
New architecture targets long prompts
Qwen 4 debuts the revamped architecture of the Qwen4 series, The Decoder reported. Instead of relying only on standard attention layers, Alibaba combines two methods. Three layers compress prior context into a condensed state, and a fourth layer then selects only the parts that matter. The approach reduces compute costs on long inputs because the model no longer applies equal attention to every part of the context.
The architecture also adds four parallel connections between model layers, with dynamic gates intended to stabilize the flow of information. An N-gram embedding layer with 328 billion parameters supplies common token sequences as ready-made building blocks. SiliconANGLE described the N-gram parameters as functioning like a cache that saves repeated calculations and removes the need to compute them from scratch.
Two attention customizations stand out, SiliconANGLE reported. Gated DeltaNet lowers the amount of hardware needed to process large prompts, while QSA skips pieces of data the model does not need to answer a prompt. Both changes are intended to reduce infrastructure costs. Alibaba said training the 2.4-trillion-parameter configuration ran stably, with training loss falling steadily and no sudden spikes, according to The Decoder.
Live Delivery and physical lab demonstrations
Alibaba calls its approach Live Delivery, in which the model delivers a usable first version early and builds on it in later passes, The Decoder reported. Users can step in and change direction at any point. The company has open-sourced the accompanying Qwen-Live-Delivery-Plugin, which SiliconANGLE described as an extension that helps Qwen 4 Max handle time-consuming tasks by delivering an initial draft and then adding improvements iteratively.
Alibaba also presented the model working in physical laboratories. Paired with Mira, a system from startup Deep Principle, the model ran autonomously for 48 hours, designed, synthesized and tested a new coolant for high-performance data center chips, searched 2.7 million molecular candidates and completed four full cycles of design, synthesis and measurement, according to Alibaba. These demonstrations illustrate the direction of travel for agentic models, although they are company claims about selected deployments.
Closed weights remain the central constraint
The largest limitation for the developer community is that Qwen 4 Max is closed-weight, Decrypt reported. Developers can access it through an API or Alibaba’s applications, but they cannot download its internal parameters, modify it or run it privately on their own hardware. That distinction matters in a market where local deployment and customization often determine whether a model can enter sensitive enterprise workflows.
Qwen has nevertheless built a substantial open ecosystem. An August report from Hugging Face counted 151,448 Qwen-based derivative models on its platform, about 4.7 times the number built specifically on Meta’s Llama family, Decrypt reported. Alibaba says smaller Qwen 4 models will be released with open weights, giving developers a route to build on this generation, though the company has not announced their sizes or release dates.
Alibaba released Qwen 3.8 Max in August, which also has 2.4 trillion parameters but activates 95 billion per token, The Decoder reported. The new Qwen 4 Max therefore lowers active parameters per token by 24 billion while raising the context ceiling and adding native image processing. With Agent Arena and Code Arena results already showing Chinese leadership in specific categories, and pricing that remains competitive, the release puts Qwen 4 Max in a credible position. Whether it can convert that position into durable adoption will depend on the open-weight releases Alibaba has promised and on how independent evaluations of its agent claims develop over time.

