Wow... this is amazing. Qwen 3.8 Flash Next running locally on a single RTX 6000 Pro.
The recipe is based on a fixed fork of SGLang, which serves RadixArk/Qwen3.8-Flash-Next-NVFP4 on a single 96GB RTX PRO 6000 and I have limited it to 275W; driver NVIDIA 610.57.04, CUDA
Perplexity just moved their whole AI agent onto your own machine.
Browsing, reasoning, tools, file access. All local. They didn't shrink the cloud version down. They rebuilt the architecture for what a local model does well. Smaller prompts, tools that only load when needed,
Alibaba's Qwen 3.8-27B is free to use right now, and it reads text, images, and video.
No GPU. No downloading 55GB of files. You go to Token Harbor, find the free route, and send a prompt. The context window is 262,000 tokens natively. Alibaba's hosted version stretches to 1
Qwen 3.8 Flash Next holds 125 billion parameters but only fires 6 billion every time it thinks.
Alibaba says it cost roughly a tenth of what their last flagship cost to train. Here's why that works. Most models read the whole book from page one every time you ask about chapter
You have reached the end of the archive
All of qwen38