Qwen3.8-2.4T-A95B: Features, Cost and Uses

TLDR

Alibaba just released Qwen3.8, its smartest open AI model yet, which is great at coding and finishing complex, multi-step tasks.
It uses a "Mixture of Experts" design, meaning it has a massive 2.4 trillion parameters but only powers up 95 billion at a time to save computing effort.
It can handle huge amounts of text at once—up to 262,144 tokens natively, and can be stretched to over 1 million.
It is a huge upgrade over the older Qwen3.7-Max, especially in writing and fixing code.
Qwen3.8-2.4T-A95B is a brand-new, heavy-duty AI language model made by Alibaba. It is built for developers, programmers, and researchers who need an AI that doesn't just answer quick questions, but can actually plan out and finish long, complicated projects on its own.

Coding and Agents: It is really good at writing code and acting as an "agent"—meaning it can plan out a multi-step task, interact with its environment, and actually finish the job without giving up halfway.
Flexible Thinking: You can control how hard it thinks. You can tune its "reasoning effort" to make it think faster or deeper, and it can remember its own thought process from previous messages to stay on track.
Developer-Friendly: It is designed to work easily with the tools and software setups that programmers already use every day.

Versus the older Qwen3.7-Max: The new Qwen3.8-Max absolutely crushes the older Qwen3.7-Max in coding tests. For example, on a coding test called "FrontierSWE," it scored 73.5 compared to the older model's 40.7. It also jumped from 21.6 to 56.6 on the "DeepSWE" test.
Versus rivals: It holds its own against top-tier competitors like Opus 4.8 and GPT 5.6 Sol. It actually beat both of them on the "SWE-bench Pro" coding test, scoring 67.7, and beat Opus 4.8 on the "FrontierSWE" test.

Context Window: It can natively handle 262,144 tokens at once (a token is roughly a piece of a word, so this equals about 200,000 words—enough to read a whole book). It can even be extended up to 1,010,000 tokens.
Architecture: It has 2.4 trillion total parameters (the "brain connections" of the AI), but only activates 95 billion at any given moment. This makes it smarter without slowing down your computer too much.
Cost: The notes do not mention exact prices, but the raw model is open-source. Alibaba also offers a paid, managed cloud version if you don't want to run it yourself.

Different Versions: The open-source model (Qwen3.8-2.4T-A95B) is just the text model. If you want extra features like vision (image input) or a built-in 1-million-token context window, you have to use Alibaba's official cloud version, called Qwen3.8-Max.
Hardware Needs: Because it is a massive 2.4-trillion-parameter model, running it yourself requires some serious, high-end computer hardware. Most regular users will need to use the cloud version instead of downloading it.

Qwen3.8-2.4T-A95B, Explained Simply