Alibaba Launches Trillion-Parameter Sparse MoE Model Qwen3.8-Max and Begins Public Beta of Enterprise Agent

On August 3, Alibaba officially released its next-generation foundation model Qwen3.8-Max. The model has a total of 2.4 trillion parameters, uses a sparse mixture-of-experts (MoE) architecture, activates approximately 95 billion parameters per inference, supports a context window of up to 1 million tokens, and natively possesses multimodal vision capabilities. Built on the Qwen 3.5 architecture and jointly optimized through sparse MoE and mixed attention mechanisms, it has improved performance in coding, professional office tasks, scientific research, and long-cycle tasks. The Qwen3.8 API is now available on the Qwen AI platform for global developers.

On pricing, domestic rates are RMB 12 per million input tokens, RMB 36 per million output tokens, and RMB 1.5 for implicit cache hits; international rates are $2 per million input tokens, $6 per million output tokens, and $0.25 for implicit cache hits.

In third-party benchmarks, Qwen3.8-Max delivered strong results across several metrics. On the authoritative Arena leaderboard, the model ranked behind Anthropic's Claude series. On CodeArena, Qwen3.8 placed fourth globally; on Vision Arena, it ranked second globally. Specific evaluation scores include: PaperBench 93.0, instruction following IF Bench 82.8, scientific reasoning GPQA Diamond 92.6, visual reasoning BabyVision 82.0, and OSWorld-Verified 86.1.

For long-cycle complex tasks, Qwen3.8-Max demonstrated autonomous execution capability. In an automated coding test, the model ran independently for 16 days without human intervention, completing the automatic construction and self-evolution framework for the oh-my-cli project and generating 265 code commits on GitHub. In a scientific research scenario, the model worked autonomously for about 125 hours, successfully reproduced the relevant paper and completed four rounds of self-evolution exploration. In a chip design sandbox, after about 500 rounds of interaction, the model reduced the hardware gate count from 8,298 to 678. In a 365-day long-cycle e-commerce business simulation benchmark, the model took first place with total capital of RMB416,252.

On the same day, Alibaba's enterprise agent product QwenWork opened for public beta and integrated Qwen3.8. The product, formed from the consolidation of three existing products, has initially connected to DingTalk IM and will later connect to enterprise databases and workflows. During the beta, users can access the web version and standalone PC client through the QwenWork official website, and entry points on DingTalk PC and mobile versions will also open. In addition, Alibaba said it will launch an international version of QwenWork.

In the competitive landscape of domestic large models, five major Chinese large models have all completed flagship iterations within less than 100 days. Prior to this, MiniMax released its next-generation multimodal generative model MiniMax H3, ByteDance launched the video creation model Seedance 2.5, and DeepSeek released the V4 Flash model. Following the release of Qwen3.8, Alibaba's Hong Kong shares closed up 7.01% at HK$125.2 per share, pushing its market capitalization back to HK$2.4 trillion; its US-listed shares closed up about 4.5%.