Alibaba Releases Latest AI Model Qwen3.8-Max with 2.4 Trillion Parameters, Reaching World-Class Level

Alibaba has released its latest large AI model, Qwen3.8-Max, featuring a total of 2.4 trillion parameters. It is the most powerful model in the Qwen family to date and will be open-sourced next week. According to a report by *IT Home*, in the third-party model evaluation leaderboard Arena released on the 3rd, Alibaba’s Qwen model ranked second only to Anthropic’s Claude series, placing its overall performance at the global top tier among large models.

According to the “Qwen Large Model” WeChat official account, Qwen3.8-Max is built on the Qwen3.5 architecture, with its parameter scale expanded to 2.4 trillion and 95 billion activated parameters. It supports a context window of 1 million tokens and delivers comprehensive improvements in programming, office work, scientific research, and long-horizon tasks. The model can handle more challenging problems and more reliably complete complex end-to-end tasks, producing trustworthy results. The model weights will be open-sourced next week, along with the Qwen3.8-27B model.

Currently, developers worldwide can access Qwen3.8 API services through the Qwen AI platform. Pricing in mainland China is 12 yuan per million input tokens and 36 yuan per million output tokens, with implicit cache hits at 1.5 yuan. Overseas pricing is $2 per million input tokens and $6 per million output tokens, with implicit cache hits at $0.25. Compared with similarly capable frontier models available overseas, Qwen3.8 offers higher cost-effectiveness.

The Qwen team highlighted the model’s programming capabilities. In one test, Qwen3.8-Max was required to create a project from scratch and build a self-evolving harness through long-horizon autonomous coding over a period of about ten days. Results showed that as of July 30, 2026, after approximately 16 days of fully AI-driven operation, the repository accumulated 265 code commits, 127 merge requests, and 151 issue tickets, demonstrating continuous evolutionary autonomous programming ability.

In professional office scenarios, Alibaba stress-tested Qwen3.8-Max’s ability to deliver production-grade results across high-frequency work scenarios drawn from hundreds of high-value occupations. For example, acting as a corporate lawyer, the model reviewed a corpus of more than 100 documents in a single pass, identifying 1,284 relevant clauses and completing the full review within one hour. In another case, as a structural engineer, it reconstructed a 30-story office building’s seismic structural model in a browser using only a single blueprint.

For long-horizon tasks, Alibaba constructed the E-Commerce Bench based on real desensitized transaction data from Taobao and Tmall, simulating 365 days of e-commerce operations. Qwen3.8-Max was given 100,000 yuan in starting capital to simultaneously operate multiple online stores, independently handling product selection, supply-chain negotiations, and other tasks. The model ultimately achieved a total cash balance of 416,000 yuan, outperforming the second-place GLM 5.2 by 38% and improving 152% over the previous flagship model Qwen3.7-Max.

IT Home further reported that in the Arena leaderboard published on the 3rd, Alibaba’s Qwen models ranked just behind Anthropic’s Claude series and were assessed as reaching the global top level in overall performance. The Qwen3.8 API is already available on the Qwen AI platform and has been integrated into Alibaba’s newly launched AI agent product “Qwen Office” released on the same day.