Key Highlights
- Qwen 3.8-Max launched with 2.4 trillion parameters and 95 billion active parameters.
- Alibaba plans to release Qwen 3.8-Max open weights next week.
- Qwen3.8-27B will also become available as an open-weight model.
- The model targets coding, research, workplace productivity and long-horizon reasoning.
- Alibaba says the model supports autonomous multi-day software development workflows.
- Native multimodal reasoning integrates vision into planning, execution and self-correction.
- API pricing starts at $2 per million input tokens and $6 per million output tokens.
- Qwen 3.8-Max is available through QwenCloud and Qwen Studio.
Alibaba’s Qwen team has introduced Qwen 3.8-Max, describing it as the most capable model in the Qwen family to date. The company also announced that the open weights of Qwen 3.8-Max will be released next week, while Qwen3.8-27B will also become available as an open-weight model.
According to the company, Qwen 3.8-Max is built on the architectural foundation of Qwen 3.5 and scales to 2.4 trillion parameters, with 95 billion active parameters. The model is designed to improve performance across coding, research, workplace productivity, and long-horizon reasoning tasks, while supporting end-to-end execution of complex workflows.
Alibaba said Qwen 3.8-Max is now available through QwenCloud APIs and Qwen Studio. The company has positioned the model as its flagship AI offering for developers and enterprises seeking advanced coding, reasoning, and multimodal capabilities.
The company highlighted autonomous software engineering as one of the model’s primary strengths. It said Qwen 3.8-Max can independently handle multi-day coding projects, progressing from an empty project directory to production-ready software without continuous human intervention. Alibaba stated that the model continuously improves through feedback loops, enabling it to refine its own development process while completing complex engineering tasks.
Beyond coding, Alibaba said the model is intended for production-quality work across multiple industries and professional use cases. It also supports long-horizon planning, allowing it to execute extended workflows involving hundreds of reasoning steps. The company cited examples including chip design optimization over more than 500 iterations and year-long e-commerce strategy planning.
Qwen 3.8-Max also introduces native multimodal capabilities that integrate visual understanding into the reasoning process. Rather than treating images solely as inputs, the company said the model incorporates visual information into planning, execution, and self-correction throughout task completion.
Alibaba shared benchmark results comparing Qwen 3.8-Max with several leading AI models across software engineering, reasoning, web development, research reproduction, and perception tasks. The published charts indicate that the model achieved leading or competitive scores on benchmarks including SWE-Pro, TerminalBench, PaperBench, FrontierSWE, QwenReactBench, CoWorkBench, JobBench, Agents’ Last Exam, BabyVision, CharXiv, ERQA, LVBench, Vision2Web, MobileWorld, and OSWorld-Verified.
The company also announced pricing for API access. Input tokens are priced at $2 per million tokens, while output tokens cost $6 per million tokens. Implicit caching is priced at $0.25 per million tokens.
Alibaba said developers can begin using Qwen 3.8-Max immediately through QwenCloud, while the release of the model’s open weights is scheduled for next week, marking the first time the company plans to open-source the weights of a Qwen Max-class model.
