Alibaba drops Qwen3.8-Max with 2.4T parameters

Imagine handing an AI a running software app with no source code and no internet connection, and telling it to figure out how it works and rebuild it from scratch.

That used to sound like science fiction, but Alibaba just made it a benchmark. The company officially unveiled Qwen3.8-Max, its largest and most capable flagship model to date.

Built on an MoE architecture with hybrid attention, the model packs a massive 2.4 trillion total parameters while activating just 95 billion during inference. It boasts a 1 million token context window and already ranks fifth in Text Arena, second in Vision Arena, and fourth in Frontend Code Arena.

A 16-day coding marathon without humans

Where Qwen3.8-Max really turns heads is in autonomous execution across long horizons. In an internal test, the model ran a real-world software engineering project for 16 consecutive days without any human intervention.

It built a self-evolving agent framework from scratch by processing user feedback, running its own test suites, and analyzing logs. The resulting project, oh-my-cli, is now fully open-sourced on GitHub.

Alibaba also introduced RecreationBench, a benchmark where the model explores closed-source apps purely through real-time visual feedback and black-box interaction, rebuilding them from the ground up.

Beyond coding, it outperformed humans in the WWW2025 Multimodal Dialogue Intent Recognition Challenge. It can digest 100-hour livestreams, turn UI screenshots into working frontend code, and generate 3D interiors from 2D floor plans.

The model is available via API on Alibaba Cloud Model Studio and inside Alibaba's new workplace agent platform, QwenWork. The biggest news? Alibaba plans to open-source the model weights next week.

Why it matters

We think the real shockwave here is Alibaba's decision to open-source the weights of a 2.4T parameter flagship.

Nearly every model sitting in the top tier of the Arena leaderboards is locked behind closed APIs. Dropping open weights for a system that ranks second in vision and fifth in text completely resets expectations for what open-weight AI can deliver.

For enterprise developers, the appeal lies in autonomous problem-solving. Running unattended for 16 days proves that AI agents are graduating from minor script assistants to autonomous project contributors.

That said, running this monster won't be cheap. Activating 95 billion parameters helps inference latency, but holding 2.4 trillion parameters in memory is an infrastructure nightmare for anyone without high-end clusters. We think enterprise cloud adoption will surge immediately, while on-prem deployments will remain reserved for well-funded labs.