- The 2.4 trillion-parameter model ranks among the world’s top AI systems and is designed to handle coding, professional tasks and multimodal workflows
- By open-sourcing its flagship model, Alibaba is stepping up the race for developers and enterprise AI adoption
Alibaba Group on August 3 unveiled Qwen3.8, the largest and most capable foundation model in its Qwen series, as Chinese AI developers intensify competition to build models that can perform increasingly complex tasks with less human supervision.
With 2.4 trillion parameters, Qwen3.8 ranks among the world’s largest publicly available AI models. On the Arena benchmark, a widely followed third-party leaderboard, Qwen models rank behind only Anthropic’s Claude series in overall performance, placing Alibaba among the leading global AI model developers.
The biggest upgrade comes in coding capabilities. Qwen3.8 is moving beyond the role of a programming assistant that completes snippets of code and toward functioning as an autonomous software developer.
On CodeArena, a benchmark focused on programming performance, the model ranked fourth globally.
While earlier-generation AI systems could generate functions or assist with debugging, Qwen3.8 can reportedly start from an empty folder, independently write code, build frameworks, run tests and fix errors over an extended period to create a functional agent framework without human intervention. The project has been fully open-sourced.
Long-horizon tasks
Beyond coding, Qwen3.8 is designed for professional workflows across hundreds of job scenarios, producing deliverables that can be directly used in business settings.
Alibaba said the model can significantly shorten complex knowledge tasks. For example, a legal team that may spend a week reviewing and labeling hundreds of legal documents could complete the same process within about an hour using Qwen3.8.
The model also features long-horizon task capabilities, allowing it to plan, adjust strategies and improve performance through thousands of rounds of interaction.
Qwen3.8 supports multimodal understanding, including the ability to analyze 200-page financial reports, process more than 100 hours of video content and integrate visual information into workflows.
The API version is now available through Alibaba’s Qwen AI platform. The company said its international input and output pricing is about 40% and 24% of Anthropic’s Opus5 pricing, respectively. Qwen3.8-Max is expected to be open-sourced next week.
From smarter answers to autonomous execution
Qwen3.8 arrives as the AI industry moves from conversational models toward agentic systems that can independently complete multi-step tasks.
The ability for AI systems to build software, analyze documents and execute workflows is pushing competition beyond model size and benchmark scores toward real-world productivity.
For global developers, Alibaba’s open-source strategy is also a play for ecosystem influence. By making a trillion-parameter-scale model available to the wider community, Alibaba aims to attract developers and companies that can build applications on top of Qwen.
The resulting feedback loop — wider adoption generating more usage data and engineering improvements — is becoming a key battleground in the global AI race.
As foundation models evolve from tools that answer questions into systems that perform work, control of the developer ecosystem may prove as important as raw model capability.
