Alibaba Debuts 2.4 Trillion Parameter Qwen3.8-Max AI Model, Stock Jumps Over 7%

On August 3, Alibaba Group officially released the new-generation base large model Qwen3.8-Max. The model contains 2.4 trillion total parameters with 95 billion activated parameters, making it the largest and most powerful flagship in the Qwen series. Built on the Qwen3.5 architecture, it uses a sparse mixture-of-experts structure jointly optimized with a hybrid attention mechanism, supports a context window of up to 1 million tokens, and natively enables visual understanding. The model delivers improved capabilities in programming, professional office tasks, scientific research and long-cycle assignments, allowing end-to-end completion of complex tasks with limited human intervention.

The news lifted Alibaba’s Hong Kong-listed shares, which closed at HK$125.2, a 7.01% gain for the day on turnover of HK$23.9 billion, propelling total market capitalisation back to HK$2.4 trillion. In the US, Alibaba shares rose more than 5% to US$127.36. From the phase low of HK$88.65 touched in late June, the Hong Kong shares have rebounded more than 41%.

On the independent third-party Arena leaderboard, Qwen3.8-Max ranked fifth in Text Arena and second in Vision Arena, placing the model in the top echelon of global large models. It placed fourth worldwide in the coding agent benchmark CodeArena. On PaperBench, which measures paper comprehension, it scored 93.0, ahead of Anthropic’s Fable 5 at 88.8; on instruction-following IFBench, it achieved 82.8 versus Fable 5’s 63.5. On visual reasoning BabyVision without tools, it scored 82.0; on OSWorld-Verified, which evaluates an agent’s ability to operate a computer, it led mainstream models with 86.1.

In an automated programming test, Qwen3.8-Max ran unsupervised for 16 days, autonomously constructing and evolving the oh-my-cli project and generating 265 code commits on GitHub. In a scientific research scenario, the model worked continuously for about 125 hours, successfully reproducing an AI Reasoning paper and conducting four self-evolution cycles, improving the AIME24 benchmark score by 2.7 points. In the WWW2025 multimodal dialogue intent recognition challenge on Alibaba Cloud’s Tianchi platform, the model beat 458 human teams within 24 hours. Inside a chip design sandbox, after roughly 500 rounds of interaction, it reduced the hardware gate count from 8,298 to 678. On the 365-day long-cycle e-commerce simulation benchmark E-Commerce Bench, the model won first place with total capital of RMB416,252.

The Qwen3.8-Max API is now available on the Qwen AI platform for developers worldwide. Domestic pricing is RMB12 per million input tokens, RMB36 per million output tokens, and RMB1.5 for implicit cache hits. Overseas pricing is US$2 per million input tokens, US$6 per million output tokens, and US$0.25 for implicit cache hits.

On the same day, Alibaba launched an enterprise-grade agent product named “Qwen Office,” consolidating QoderWork, MuleRun and Wukong. The product runs on Qwen3.8 and has entered public beta. It has already achieved basic integration with DingTalk IM, with plans to connect to enterprise databases and workflows, launch an independent mobile app, and release an international version.