Alibaba has officially launched Qwen3.8-Max, a new artificial intelligence model featuring 2.4 trillion parameters. The company stated this is its most capable model yet, built upon the Qwen3.5 architecture. Approximately 95 billion parameters are active during inference. Alibaba claims performance comparable to leading models from U.S. firms such as Anthropic and OpenAI. The company also announced that open-source weights for Qwen3.8-Max will be released within the week, a first for a model in the Qwen-Max class.
The model supports a context window of up to 1 million tokens, allowing it to process extensive documents, code repositories, and long videos. Alibaba highlighted its capabilities in software engineering, professional work, research, and multimodal tasks. In demonstrations, Qwen3.8-Max autonomously developed a command-line tool project called "oh-my-cli" over 16 days, accumulating 265 commits and 151 issues. The model also reproduced and improved upon the findings of a research paper on data selection for AI training, writing approximately 7,600 lines of code and conducting 33 GPU training rounds. This iterative process led to a 2.7 percentage point improvement on the AIME24 math benchmark compared to the paper's original approach.
Alibaba also entered Qwen3.8-Max into the WWW2025 Multimodal Dialogue Intent Recognition Challenge on its Tianchi platform. Competing against 526 human teams under a 24-hour limit, the model developed a system combining language, vision-language, and image models, achieving an accuracy of 0.853 and outperforming 87% of participants.
For multimodal tasks, Qwen3.8-Max can analyze financial reports and PDFs exceeding 200 pages, and organize videos longer than 100 hours into searchable structures. The model produced outputs such as websites, animations, and interactive applications. In a simulated year-long e-commerce environment, Qwen3.8-Max managed product selection, negotiations, inventory, pricing, and capital allocation, reportedly finishing with a balance of ¥416,252, a 4.16-times return. This result was 38% higher than the second-place model and 152% higher than the previous Qwen3.7-Max generation.
Benchmark results published by Alibaba show Qwen3.8-Max scoring 93.0 on PaperBench and 82.8 on IFBench. On Terminal-Bench 2.1, it achieved 86.6, placing it behind GPT-5.6 Sol (max) at 88.8, but ahead of Claude Opus 4.8 and Claude Fable 5. Multimodal benchmarks include scores of 86.1 on OSWorld-Verified and 92.1 on OmniDocBench 1.5. Alibaba stated that some of these tests were conducted internally, and evaluation methods varied across models. Independent verification of these results is pending.
Alibaba is also releasing a second checkpoint, Qwen3.8-27B, which is suitable for standard on-premise GPU hardware, unlike the larger 2.4 trillion-parameter flagship model which is described as a datacenter artifact. The company plans to release the weights for both models through platforms like Hugging Face and ModelScope. Qwen3.8-Max is accessible via API through QwenCloud.