Alibaba Qwen Qwen3.8 Official Announcement: 2.4T Parameters Coming Soon to Open Source, Preview Version Now Live

Alibaba is making another big move. Yesterday, Alibaba's Qwen suddenly announced something: Qwen3.8 is here, with 2.4T parameters, and it's about to be open-sourced. You read that right—2.4 trillion. And the official statement itself says this might be the strongest model out there aside from Fable 5. That's a bold claim, but Alibaba isn't the first to say it. Last week, Kimi K3 just set a new open-source record with 2.8T parameters, and this week Alibaba

💡 What You Will Learn

Alibaba is making another big move. Yesterday, Alibaba's Qwen suddenly announced something: Qwen3.8 is here, with 2.4T parameters, and it's about to be open-sourced. You read that right—2.4 trillion.

📜 Table of Contents

• Alibaba Drops Another Bombshell
• How Big Is 2.4T Parameters, Really?
• You Can Try It Right Now, and It's Ridiculously Cheap
• Open-Source LLMs Are "Going at It" This Week
• What Can Regular People Use It For?

Alibaba Drops Another Bombshell

Yesterday, Alibaba's Qwen suddenly made a big announcement: Qwen3.8 is here, 2.4T parameters, open source coming soon.

You read that right — 2.4 trillion. And the official word is that this could be the strongest model out there besides Fable 5. Bold claim, but Alibaba isn't the first to talk big. Last week, Kimi K3 just set a new open-source record with 2.8T parameters, and this week Alibaba pulls out Qwen3.8 — two domestic large models, less than a week apart, dropping bombs back-to-back. What's even crazier is that the Qwen3.8-Max preview went live yesterday, so you can use it right now.

From Qwen3.5 to 3.6, then 3.7, Qwen's iteration speed has been dizzying. This time it jumps straight to 3.8, with parameters surging to 2.4T — a sign that Alibaba has found a new path in ultra-large model training.

How Big Is 2.4T Parameters, Really?

A lot of people don't have a frame of reference for "2.4T." Here's the deal: DeepSeek V4 Pro's total parameters are 1.6T, already top-tier among open-source models. Qwen3.8 beats it by 50%.

Bigger parameters don't just mean "bigger." The upgrade Alibaba is emphasizing this time is complex long-horizon tasks — full-stack development, data analysis, multi-step Office workflows. In other words, Qwen3.8 isn't the kind of model that answers one question at a time; it's the kind that helps you complete an entire task from start to finish.

For example, if you ask it to "build a blog system with a database," it won't just hand you a snippet of code — it'll understand the whole project structure, from requirements analysis to frontend and backend implementation, all in one go. This kind of capability has rarely been seen in open-source models before. And reportedly, in internal testing, Qwen3.8-Max-Preview's performance on long-horizon tasks is already approaching the level of world-leading closed-source models — a significant breakthrough for open source.

You Can Try It Right Now, and It's Ridiculously Cheap

The Qwen3.8-Max-Preview is already live on three platforms: Token Plan, Qoder, and QoderWork. The Qwen PC client also lets you select it directly.

Here's the kicker — the price.

Token Plan's personal tier starts at just 39 RMB/month, with a limited-time offer: Credits are discounted to 10% during the day, and as low as 2% at night (10 PM to 8 AM). Do the math — using it at night to write code or run analysis might cost you less than a cup of milk tea. The Qwen PC client is even more straightforward — just check it in the model list and use it, completely free.

For developers, the Qoder platform provides a full development environment, letting you write code and run tests in the cloud without configuring your own setup. That's super convenient for anyone who wants to quickly experience what the new model can do — no deployment hassle, just open a browser and you're using a 2.4T-parameter model.

Also, the Qwen PC client has integrated the preview version — just toggle it in the model list, similar to the ChatGPT experience. It's the easiest way for regular users to get started.

Open-Source LLMs Are "Going at It" This Week

If you follow domestic large models, you might be a bit overwhelmed this week:

• Kimi K3 (2.8T parameters) just went open source, and even Musk chimed in calling it "impressive"
• Qwen3.8 (2.4T parameters) was announced yesterday, and the preview is testable today
• DeepSeek V4 official release is reportedly dropping tomorrow

Three domestic large models, all within one week, firing on all cylinders. Who would've thought that a year ago? And all three are emphasizing long-horizon tasks and Agent capabilities — it's not about who's smarter in single-turn conversations, but who can make the model complete an entire job in one go. The competition in this lane has already escalated from a "parameter race" to a real-world test of "can it actually get things done."

One more thing worth noting: MiniMax's next-gen model, M3 Pro, is also about to launch, with parameter scale between 2.5 trillion and 3 trillion, also planned for open source. That means the competition among domestic open-source models in the second half of the year is only going to get fiercer. For regular users and developers, options keep growing, but the hardware bar for running these models keeps rising too — that's the "happy problem."

What Can Regular People Use It For?

The Qwen series has always been hugely popular on Hugging Face, with countless quantized versions popping up, and it's one of the top choices for local AI deployment in China. Once Qwen3.8 goes open source, expect community experts to quickly cook up 4-bit and 8-bit quantized versions.

So if you're planning local deployment, it's best to wait for the quantized versions before testing. 2.4T raw parameters aren't something a regular computer can handle, but quantizing it down to a usable size is just a matter of time. Plus, the Qwen community is very active — from Qwen2.5 to the Qwen3 series, every generation comes with tons of quantization tutorials, deployment guides, and use cases, so the barrier to entry is actually pretty low.

Of course, if you can't wait, trying the preview via Token Plan or the Qwen PC client costs next to nothing — 10% during the day, as low as 2% at night. A cup of milk tea's worth of money can run it all night.

Will you be giving the Qwen3.8 preview a try?