<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
        <title>it shouldn't be able to run models this big</title>
        <link>https://tube.blueben.net/videos/watch/7547b476-4d2e-4b0a-bc66-8d0bf6b0459b</link>
        <description>Get $600 off either edition of WiCi One, the first wireless GPU, using my link: https://ntck.co/wici_gpu I ran DeepSeek V4.1 Flash, a 750 billion parameter model that normally needs over 500 GB of memory, on a box with only 32 GB of VRAM. Over Wi-Fi. By the numbers that shouldn't work. The box is WiCi One, a GPU every device in your house can share. In this video you'll learn why big local models don't fit on consumer GPUs, and how WiCi got one running anyway. ☕☕ COFFEE and MERCH: https://ntck.co/coffee *Sponsored by WiCi RESOURCES / LINKS: 🖥️ WiCi One (RTX 5060 Ti edition, or the RTX 5090 edition I tested), $600 off with my link: https://ntck.co/wici_gpu 🧠 InferSSD, WiCi's open-source engine for running models bigger than your VRAM off an SSD: https://github.com/wici-ai/inferssd 🤖 DeepSeek V4.1 Flash on Hugging Face: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash 🎨 Lovart (AI image and video generation): https://www.lovart.ai 🎨 Figma: https://www.figma.com Figma and Lovart are cloud apps. For this demo, their AI generation requests were redirected to a model running locally on WiCi One. TIMESTAMPS: 0:00 - The first Wi-Fi GPU: an RTX 5090 my laptop thinks is its own 0:38 - What WiCi One is, and what's inside the box 2:15 - Why Wi-Fi should be way too slow for a GPU 4:37 - The WiCi protocol: how they made the game and GPU stop talking so much 5:36 - From Tencent cloud gaming to a wireless GPU 6:05 - The two problems holding local AI back 7:32 - Why you can't run frontier-size AI models at home 9:57 - The math on running a 750 billion parameter model 11:59 - Mixture of experts, explained with badminton 13:50 - The four tricks that make DeepSeek V4.1 Flash usable on 32 GB of VRAM 15:16 - Running DeepSeek V4.1 Flash live on WiCi One 17:36 - Why GPUs don't like to share 19:41 - Testing WiCi's scheduler: switching AI models with and without it 21:06 - The family GPU: my verdict, pricing and shipping 22:18 - A prayer for you **Sponsored by WiCi SUPPORT NETWORKCHUCK: ☕☕ COFFEE and MERCH: https://ntck.co/coffee READY TO LEARN?? 🔥🔥Join the NetworkChuck Academy!: https://ntck.co/NCAcademy 📚 CCNA Course: https://ntck.co/ccna FOLLOW ME EVERYWHERE: Instagram: https://www.instagram.com/networkchuck/ X/Twitter: https://x.com/networkchuck Facebook: https://www.facebook.com/NetworkChuck/ Join the Discord server: https://ntck.co/discord Some links in this description are affiliate links. If you buy through them, I may earn a small commission at no extra cost to you. #sponsored #wici #localai #networkchuck</description>
        <lastBuildDate>Sat, 31 Oct 2026 15:46:41 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>PeerTube - https://tube.blueben.net</generator>
        <image>
            <title>it shouldn't be able to run models this big</title>
            <url>https://tube.blueben.net/client/assets/images/icons/icon-96x96.png</url>
            <link>https://tube.blueben.net/videos/watch/7547b476-4d2e-4b0a-bc66-8d0bf6b0459b</link>
        </image>
        <copyright>All rights reserved, unless otherwise specified in the terms specified at https://tube.blueben.net/about and potential licenses granted by each content's rightholder.</copyright>
        <atom:link href="https://tube.blueben.net/feeds/video-comments.xml?videoId=7547b476-4d2e-4b0a-bc66-8d0bf6b0459b" rel="self" type="application/rss+xml"/>
        <item>
            <title><![CDATA[it shouldn't be able to run models this big - Jeena]]></title>
            <link>https://tube.blueben.net/w/ftYh4DvdMgX5BaE5xcw2cZ;threadId=6274</link>
            <guid>https://tube.blueben.net/w/ftYh4DvdMgX5BaE5xcw2cZ;threadId=6274</guid>
            <pubDate>Fri, 30 Oct 2026 21:56:41 GMT</pubDate>
            <content:encoded><![CDATA[<p>That was a very informative video, thanks for taking us so deep down and explaining the details. It's a really interesting idea. The only thing right now is that there is hack on hack on hack to make it work, which is impressive - don't take me wrong - but it seems very fragile and you need to believe that they will keep doing the model preparations for you for each new model, etc. and that they don't go bankrupt. On top of that the price is very high just for the device and then you still need the GPU and the separate SSD for it, if you don't yet have a PC then yeah might make sense. If you already have one like you and me, I'm thinking we should be able to do all those optimizations for the PC, there was no proprietary hardware where from what I've seen from your explanation. And some parts are already happening, like with Strata, etc.</p>
]]></content:encoded>
            <dc:creator>Jeena</dc:creator>
        </item>
    </channel>
</rss>