Qwen 3.8 Flash Next, a new open-weight AI model, uses three architectural tricks to match larger rivals while running at 38 tokens per second. It outperforms DeepSeek-4 Pro and stays free to download, collapsing the gap between expensive and accessible AI.
This is the new QWEN 3.8 Flash Next AI. And the first results look fantastic. Incredible system, and we are lucky to get so many of these open and free AI systems that it's easy to confuse which is which. Fellow Scholars! That is a very good problem to have. And it seems that we can switch out paid, closed systems to open weights AI that we can download and run ourselves, Forever. No subscriptions. And we have a paper too. We are spoiled here, and I am really grateful. And that is why I keep covering as many of these as I can. I want to show you how I, and you Fellow Scholars, are already using it out there in the wild. Now, there is Qwen 3.8 Max, incredibly huge. Then, the 27 billion. This runs on a very beefy laptop or a desktop. This is 3.8 Flash Next. This is in the middle. But it is so Different. The small guy, 27B, is a dense model. Now, this one is a mixture of experts model. Very different. This means that for each token, it only uses a small part of the model. Great for systems with lots of memory and slower memory bandwidth. Like the DGX Spark. Or 2. I am running experiments with it on 2 Sparks, day one, it runs about 38 tokens per second. Very respectable. Thank you so much to Jensen, Brad, and Mark from NVIDIA for the hardware. Now, you brilliant Fellow Scholars already found a way to run this on much more modest hardware. Now, 3.8 Flash Next is the first of their next-generation model and has substantial innovations. Here's 3. Dear Fellow Scholars, this is Two Minute Papers with Dr. Károly Zsolnai-Fehér. 1. QWEN sparse attention. As we start adding documents, or just keep talking to the AI longer, its memory we call context starts filling up. And with that, things get really expensive. This is called full attention, and it has quadratic complexity. So processing twice as much context is roughly 4 times as much work. DeepSeek tried to ease this with DSA, which picks only important individual tokens. Quentn's new QSA, however, goes further. It bundles these tokens up into tiny blocks, and only searches those blocks. Why? Well, because it makes this kind of growing context even cheaper. 2, for each token, the model carries additional information about that token. But here, the layers can step on each other's toes because they all keep rewriting the same running information. Now here, instead they use not one, but 4 branches. Why? Because here, information can be left alone while the rest is changed. They call it gated residual. And 3, imagine a hot dog. Yummy! Now, the 2 words separately mean something quite different, right? Hot and dog does not equal hot dog. So, QuEN can bunch these short token combinations together and build lookup memory for them to retrieve from quickly and cheaply. They call it n-gram embedding. It's a variant of what DeepSeek already does. DeepSeek spreads this lookup memory across multiple layers, while QWEN puts it into one huge lookup layer near the start. And these 3 give you a system that already outperforms some of the best open weights AI systems, maybe even DeepSeek-4 Pro that is much, much bigger. And all this just a few days after release. This is mind-blowing. And we get all this for free forever. What a time to be alive! We need new tools for the era of LLMs, and Weights Biases now has Weave, a lightweight toolkit to confidently iterate on LLM applications. Use traces to debug how data flows through each step of your app, and use evaluations to measure your progress. It is the best. Try it out now at wnb.me/papers or click the link in the description below.
