Alibaba’s Qwen team has officially released Qwen3.8-Max, describing it as the largest and most capable model in the Qwen family to date. Built on the architectural foundation of Qwen3.5, the flagship model contains 2.4 trillion total parameters, with approximately 95 billion parameters activated during inference. It supports a context window of up to 1 million tokens, enabling it to work with large collections of documents, extensive software repositories, long videos and complex tasks that require sustained reasoning across many steps.
Qwen3.8-Max is available through QwenCloud and can be called through its API. The company also plans to release the model’s weights through Hugging Face and ModelScope, marking the first time Alibaba has offered open weights for a Qwen-Max-class model.
Alibaba said the model was developed to produce dependable, end-to-end deliverables rather than only answer individual questions. Its primary areas of improvement include software engineering, real-world professional work, research, multimodal agents and long-horizon tasks that may require hours or days of autonomous operation.
In one autonomous coding demonstration, Qwen3.8-Max created the “oh-my-cli” project from an empty repository and continued developing it for more than 10 days. The model built a self-evolving engineering system that converted user and developer feedback into issues, assigned tasks to agents, generated code and ran tests before merging approved changes.
After approximately 16 days of autonomous operation, the repository had accumulated 265 commits, 127 pull requests and 151 issues. The system included automated build, unit, end-to-end and desktop lifecycle testing, along with monitoring and recovery processes for detecting and correcting abnormal states.
In another test, the model was asked to reproduce the experiments from a research paper about selecting valuable training data and then improve on the paper’s methodology. Starting without sample code or an existing pipeline, Qwen3.8-Max reportedly worked continuously for approximately 125 hours, wrote about 7,600 lines of code and completed more than 1,100 actions and 33 rounds of GPU training.
The model first recreated the paper’s six primary findings and then spent approximately 88 hours testing 18 additional ideas through an iterative research process. Alibaba said the resulting method improved performance on the AIME24 math benchmark by 2.7 percentage points compared with the paper’s original approach.
Qwen3.8-Max was also entered into Alibaba Cloud’s WWW2025 Multimodal Dialogue Intent Recognition Challenge, where it competed against 526 human teams. Working without human assistance under a 24-hour limit, the model created a system combining multiple language, vision-language and image models.
Across 45 submissions, the system’s reported accuracy increased from 0.60 to 0.853, placing it ahead of 458 participating teams, or approximately 87% of the field. Alibaba said the demonstration showed the model’s ability to respond to evaluation feedback, revise its methods and improve its performance over repeated cycles.
The company also tested Qwen3.8-Max across workflows involving hundreds of professions. In one example, the model identified 1,284 relevant clauses across hundreds of corporate compliance documents in less than an hour. Alibaba estimated that a comparable review could require approximately one week of work from a team of paralegals.
Other demonstrations included creating an eight-screen interactive digital banking prototype without human revisions, developing a 26-dish restaurant menu from more than 100 supplier documents, reconstructing the seismic model of a 30-story office building and processing approximately 8,400 basketball possessions per player into tactical reports for coaches.
For quantitative research, Qwen3.8-Max was given a one-line description and tasked with developing an exchange-traded fund rotation strategy. The model built the data system, created investment factors, analyzed backtests and changed its approach when it detected potential overfitting.
In a larger factor-research exercise, the model divided six investment themes into 50 research directions each, dispatched approximately 330 sub-agents and completed about 6,000 backtests. Alibaba said this compressed research that might traditionally require weeks or months into an automated workflow completed during one conversation.
The model was also evaluated on autonomous semiconductor design. Starting with basic instructions and empty hardware-design templates, Qwen3.8-Max completed logic design, RTL code generation, simulation, synthesis and physical layout work for a cryptographic hardware accelerator.
During approximately 500 interaction turns and 71 evaluations, it reduced the design from 8,298 gates to 678 gates. Alibaba said the physical implementation reduced the chip area by 81%, lowered wire length from 33,369 micrometers to 4,187 micrometers and achieved timing closure at 500 megahertz.
Alibaba also tested the model in a simulated year-long e-commerce environment involving 12 store types, 60 product categories, nearly 600 suppliers and 7,000 products. Starting with ¥100,000, the model was responsible for product selection, supplier negotiations, inventory management, pricing, returns and capital allocation.
Qwen3.8-Max finished the simulation with a reported balance of ¥416,252, representing a 4.16-times return. Its balance was 38% higher than the second-place model and 152% higher than the result achieved by the previous Qwen3.7-Max generation.
The model also supports multimodal tasks involving images, documents and video. According to Alibaba, Qwen3.8-Max can analyze financial reports and PDFs longer than 200 pages, organize videos exceeding 100 hours into searchable memory structures and produce outputs such as websites, animations, three-dimensional visualizations and interactive applications.
During visual production tasks, the model can examine its intermediate outputs, identify problems involving layouts, object positions, spatial relationships and interactions, and revise its work without human intervention. Alibaba described this as a visual feedback loop spanning planning, execution, verification and iteration.
QwenCloud provides interfaces compatible with OpenAI and Anthropic API specifications, allowing Qwen3.8-Max to work with several existing coding and agent frameworks. Supported tools include Claude Code, Codex, Qoder CLI, Qwen Code and OpenClaw.
Developers can choose among three reasoning levels. The default “xhigh” setting is intended for complex tasks requiring detailed analysis, while “medium” balances accuracy and speed and “low” prioritizes faster, less expensive processing.
In benchmark results published by the Qwen team, Qwen3.8-Max scored 93.0 on PaperBench, 81.9 on WideSearch, 82.8 on IFBench, 60.2 on HealthBench and 58.3 on PRBench-Finance. Its multimodal results included scores of 82.3 on MMMU-Pro, 86.1 on OSWorld-Verified, 92.1 on OmniDocBench 1.5 and 90.4 on VideoMME with subtitles. Some tests were conducted internally by Alibaba, and evaluation methods varied across models and benchmarks.
KEY QUOTES:
“Today, we are officially releasing Qwen3.8-Max, the most capable model in the Qwen family to date. This also marks the first time we will open-source the weights of a Qwen-Max-class model.”
“Built upon the architectural foundation of Qwen3.5, Qwen3.8-Max scales to 2.4 trillion parameters, delivering comprehensive improvements across coding, work, research and long-horizon tasks. It can not only answer more challenging questions, but also complete complex tasks end-to-end with greater reliability, producing dependable deliverables.”
Qwen Team at Alibaba
