Qwen3.8’s 2.4T Parameters Are Already Live — and Open Weights Are Coming

Alibaba confirms that Qwen3.8’s 2.4T parameters are already live in preview mode. This massive model is currently accessible on Token Plan, Qoder, and QoderWork. Furthermore, an open-weight release is coming soon. This puts the model on a direct collision course with Kimi K3 for the largest freely deployable system.

The official @Alibaba_Qwen account announced this ambitious update. Unlike the closed Qwen3.7-Max, this new version heads toward open weights. Developers do not need to wait for the final drop to start building. Early access is available right now for testing.

Understanding Qwen3.8’s 2.4T Parameters Are Already Live

At 2.4 trillion parameters, this model sits in rare company. For context, Moonshot AI’s Kimi K3 is a 2.8-trillion-parameter open model. It is the only other recent open-weight contender in this tier. Therefore, Alibaba is now racing for the largest publicly accessible model title.

The team describes the system as continuously evolving. This suggests ongoing post-training refinements rather than a static release. Consequently, capabilities may shift before the full open-weight drop. Mixture-of-Experts architecture is likely at this scale. This design pattern allowed Kimi K3 to hit #1 on LMArena’s Code Arena efficiently.

Where to Access the Preview

You can access the Qwen3.8-Max-Preview through three specific channels today:

• Token Plan (International): Available via qwencloud.com for global API pricing.
• Token Plan (China): Accessible at platform.qianwenai.com for domestic users.
• Qoder and QoderWork: Developer tools that previously offered free daily calls.

This preview is an early access window, not the final release. It mirrors how previous Max-Preview versions debuted before broader availability. Engaging now gives developers a head start on understanding model behavior.

Why Open Weights Matter

The open-weight commitment makes this announcement significant beyond raw scale. Previous frontier models from this team launched as closed systems. However, Qwen3.8 breaks that pattern explicitly. A 2.4T model would dwarf earlier generations in total parameter count.

Self-hosting and fine-tuning become possible if Alibaba follows through. This has massive implications for private enterprise inference. Additionally, the competitive timing is notable. xAI’s next model is still in training. Thus, Qwen3.8 could arrive first without download restrictions.

Performance Claims and Benchmarks

The team positions this model as second only to Fable 5. Vendor benchmarks alone cannot settle such bold claims. Independent evaluations on agentic tasks will determine its true ranking. Nevertheless, the trajectory looks promising given prior performance.

More parameters and improved post-training suggest a genuine capability step up. Reasoning and agentic domains are key focus areas. Treat vendor positioning as a starting hypothesis. Community leaderboards will eventually tell the real story.

What to Watch Next

The timeline remains the key open question. “Coming soon” could mean days or weeks based on history. Monitoring GitHub and Hugging Face repositories provides the fastest signal. Licensing terms also matter significantly for adoption.

Apache 2.0 coverage would allow unrestricted commercial use. This makes the model viable for enterprise deployment without API dependency. Finally, benchmark this model if you build coding or agentic tools. The race for the largest downloadable model just gained a major contender.