Skip to main content

Alibaba Open-Sources 2.4 Trillion-Parameter Model, Challenging Top AI

In a move that's sending ripples through the AI community, Alibaba's Qwen team has just dropped a bombshell: they've open-sourced their flagship Max-level model, Qwen3.8-2.4T-A95B. This isn't just another incremental release—it's the first time the team has fully opened the doors to its most powerful model, letting developers and researchers worldwide get their hands on the weights.

So, what's the big deal? Let's break it down. The model boasts a staggering 2.4 trillion total parameters, but here's the clever part: it only activates 95 billion parameters per token. That's like having a massive library where you only pull out the books you need for each task—efficient and powerful at the same time. It natively handles an ultra-long context of 260,000 tokens, which can be stretched to a mind-boggling 1,010,000 tokens. That's enough to process entire book series or massive codebases in one go.

Under the hood, it builds on the strengths of Qwen3.5, with performance boosts across the board—whether you're coding, handling daily office work, diving into research, or relying on it for long-term autonomous agent tasks. The benchmarks are equally impressive. On PaperBench, a rigorous test for paper reproduction, and OSWorld-Verified, which evaluates computer operation skills, it ranks among the top. In some metrics, it even edges out several international mainstream models. But what really turns heads is its long-term autonomy. Imagine an AI that can independently tackle complex programming projects that span days, or autonomously reproduce and optimize research papers. That's not just a chatbot; that's a digital research assistant with serious stamina.

This open-source release is a game-changer. It democratizes access to cutting-edge AI, allowing smaller teams and independent developers to build on technology that was previously locked behind corporate walls. The implications are huge—from accelerating research to enabling new applications we haven't even thought of yet.

Of course, with great power comes great responsibility. The Qwen team has been careful to include safety measures, but the open-source nature means the community will play a big role in ensuring it's used ethically. Still, for now, the excitement is palpable. Developers are already diving in, testing its limits, and dreaming up new possibilities.

If you're curious, you can check out the model on Hugging Face or the official Qwen GitHub repo. The documentation is thorough, and the community is buzzing with tutorials and examples. Whether you're a seasoned AI researcher or just starting out, this is a golden opportunity to explore what a 2.4-trillion-parameter model can do.

As the dust settles, one thing is clear: the race for AI supremacy just got a whole lot more interesting. And with open-source giants like this, the future of AI is looking more collaborative—and more accessible—than ever.

Key Points

  • Model: Qwen3.8-2.4T-A95B, with 2.4T total parameters and 95B activated per token.
  • Context: Supports 260k tokens natively, expandable to 1M.
  • Performance: Top-tier results on PaperBench and OSWorld-Verified, rivaling international models.
  • Autonomy: Capable of multi-day programming tasks and autonomous research paper reproduction.
  • Open Source: First full open-sourcing of Alibaba's Max-level model, available to developers worldwide.