Alibaba's Qwen has officially launched Qwen4-Max, a mixture-of-experts model designed for long-duration AI agents, specialized professional tasks, and multimodal operations. It features approximately 2.4 trillion "backbone parameters," but only around 71 billion parameters are activated per token, alongside 32.8 billion N-gram embedding parameters. The model can process text, images, and videos, supports a context length of up to 1 million tokens, and is compatible with OpenAI and Anthropic API formats. In Qwen's proprietary Work360 professional task benchmark, the new model's score improved from 47.74 for Qwen3.8-Max to 52.69, while the average task cost decreased from $4.43 to $1.67. However, these results are from official internal tests; in 16 public benchmarks, Qwen4-Max secured only four first-place rankings and recorded a 6–10 record against Claude Opus 5.5, with noticeable gaps remaining in terminal programming, software replication, and chart comprehension. Currently, it is only available via cloud products and APIs, with weights not open-sourced; the company has indicated that smaller models in the Qwen4 series will be opened in the future.Article author and source: Qwen Team
2.4 trillion parameters does not mean that all 2.4 trillion parameters are executed during each inference.
Qwen4-Max employs a sparse mixture-of-experts architecture. The official figure of 2.4 trillion refers to the total number of parameters in the model's backbone; when processing each token, the router selects only approximately 71 billion parameters for computation.
The model also adds 328 billion N-gram embedding parameters to directly represent continuous phrases and local patterns. This number is not included within the 2.4 trillion backbone parameters, so simply adding all components together would result in a larger total; however, this summation cannot be directly equated to the computational cost of traditional dense models.
The architecture alternately combines gated incremental networks with sparse attention, while incorporating gated residual branches. The design goal is not to have all parameters participate in computation at all times, but rather to expand the model’s capacity to store and retrieve knowledge and behavioral patterns, while maintaining low per-token computational cost.
The million-context limit is the total upper bound for input and output budgets.
Qwen4-Max can accept text, images, and videos, and outputs primarily text, with an official context window of 1 million tokens.
When actually calling, users cannot use the full 1,000,000 tokens as input. The official documentation specifies the maximum input as approximately: 991,808 tokens in non-thinking mode and 983,616 tokens in thinking mode. Space is reserved for the model to generate responses and perform internal reasoning.
Non-thinking mode allows a maximum output of 131,072 tokens; in thinking mode, the combined total of reasoning process and final answer is approximately 262,144 tokens. In other words, "million-context" refers to the total amount of information the entire request can accommodate, not that you can input one million and then additionally output 260,000.
The model supports tool calling, structured output, and prompt caching, and is compatible with OpenAI Chat Completions, Responses API, and Anthropic Messages API. Developers can reduce interface-layer modifications when migrating existing agents, but tool formats and reasoning parameters still differ slightly across platforms.
Public reviews are strong, but not yet fully leading.
Qwen has released 16 public benchmark evaluations. Qwen4-Max achieved first place in four of them, including a score of 75.7 on the long-term software engineering test DeepSWE, 85.0 on the bioinformatics reasoning benchmark BioMysteryBench, and two professional agent evaluations.
It ranked second in four other tests. For example, AutomationBench scored 72.5, slightly below Kimi K3.1’s 73.3; APEX-Agents scored 72.1, close to Claude Opus 5.5’s 73.5.
The weaknesses are also clear. In Terminal-Bench 4.0, Qwen4-Max scored 48.5, below Kimi K3.1’s 67.2; ProgramBench scored only 31.5, while Claude Opus 5.5 achieved 65. It also did not rank in the top tier in tests such as software interface replication, general high-difficulty reasoning, chart understanding, and video indexing.
According to the official benchmark table, Qwen4-Max directly compares to Claude Opus 5.5 with a record of 6 wins to 10 losses. Therefore, a more accurate conclusion is: it is highly competitive in law, bioinformatics, and certain long-form professional agent tasks, but it is not a model that leads in all capabilities.
Different vendors may also use different prompts, inference budgets, and tool environments. Minor evaluation gaps may not reliably translate into real-world task differences; post-release unified third-party testing is more valuable for reference.
A 62% reduction in costs resulted from changes in both pricing and token usage.
In Alibaba's internal Work360 evaluation, Qwen4-Max scored 52.69, higher than Qwen3.8-Max's 47.74. Meanwhile, under maximum inference intensity, the number of tokens required to complete tasks decreased by approximately 24%; under the next lower level of high inference intensity, token usage decreased by approximately 35%.
In Terminal-Bench 4.0, token usage has decreased by nearly half. Alibaba explains that the new model can more quickly select the correct tools and solution paths, reducing ineffective searches, redundant file reads, and repeated modifications.
QwenCloud's announced international API pricing is $1.50 per million input tokens and $4.50 per million output tokens, with cached reads at $0.075. After combining lower unit prices and fewer tokens, the official calculation reduces Work360's average cost per task from $4.43 to $1.67, a decrease of approximately 62%.
This figure should not be interpreted as a fixed 62% cost reduction for every type of request. The cost structure may differ significantly for short Q&A tasks, tasks that cannot leverage caching, extremely long video inputs, or complex tasks requiring extensive reasoning output. Additionally, Work360 is an evaluation designed and operated by Qwen, and independent teams have yet to reproduce these cost figures.
The officially demonstrated Agent has begun delivering work results directly.
Qwen positions Qwen4-Max as "built for work," showcasing use cases in manufacturing, law, healthcare, research, gaming, and 3D design.
In the manufacturing case, the agent reads part drawings and generates machining programs, reducing a task that traditionally took about 475 minutes to just 24 minutes; in the legal case, it organized 198 disorganized pre-listing documents into an 86-page due diligence report.
In the medical demonstration, the model analyzed 1,433 original images from critical patient records and generated four care recommendations. The scientific demonstration included continuously running for 48 hours to design candidate coolant molecules and calling tools 454 times to screen Alzheimer’s drug candidates.
The model can also generate installable Android racing games from natural language requirements and operate Blender, Unreal Engine, Unity, and Godot to create 3D assets or interactive scenes.
These examples illustrate that Qwen is advancing models from simply answering questions to entire workflows involving reading materials, operating specialized software, and delivering files. However, all are officially selected demonstrations and do not equate to large-scale success rate testing. Medical advice cannot replace clinical validation, and chemical or drug candidates must be confirmed through experimentation—they cannot be directly labeled as new materials or new drugs based solely on computational results.
This is a cloud-based proprietary model.
Although the Qwen series is known for its open-weight models, Qwen4-Max does not provide downloadable weights, nor has it released parameter files, complete training data, or training code.
Users currently need to access via QwenCloud, Alibaba Cloud Bailing, Qwen Studio, QwenWork, or Qoder. The official statement indicates that smaller models from the Qwen4 series will eventually be introduced to the open ecosystem, but no specific details regarding size, licensing, or release timeline have been disclosed.
Therefore, Qwen4-Max is more similar to the flagship proprietary model sold via API and cannot be assumed to support local deployment or free fine-tuning simply because it belongs to the Qwen series.



