Artificial intelligence usually requires massive, expensive infrastructure to run. Tech giants like Google and Meta routinely spend over $1 billion building warehouse-sized data centers packed with high-end processors. For individual engineers and hobbyists, running a top-tier large language model (LLM) at home has always felt like a distant dream. However, a tech enthusiast recently shattered that barrier by running a massive 1-trillion-parameter model right on a normal home workstation, using cheap, second-hand server parts.
An anonymous builder going by the handle APFrisco shared their incredible hardware experiment on the popular Reddit forum r/LocalLLaMA. The builder managed to get Kimi K2.5—a state-of-the-art, open-source model with a massive 1-trillion-parameter architecture—running locally. Even more surprising, the system generated text at roughly 4 tokens per second. While that speed might not win any races against cloud-based supercomputers, it is fast enough to read comfortably, proving that you do not need a fortune to run frontier-class AI.
The secret behind this technical milestone lies in memory capacity rather than raw computing power. Most local AI developers get stuck because massive models require incredible amounts of memory. If the model does not fit into the computer’s graphics card memory (VRAM) or system RAM, performance collapses into a slow crawl. To solve this bottleneck, the builder sourced six discontinued Intel Optane Persistent Memory modules, giving the system a massive 768 gigabytes of memory.
Intel originally designed these Optane modules to sit in a strange middle ground between traditional system memory (DRAM) and fast storage drives (SSDs). Although Intel eventually killed off the product line and took a heavy write-off on the technology, these memory sticks have flooded the second-hand market. The Reddit builder bought the used 128-gigabyte modules for much less than what the equivalent capacity of standard DDR4 DRAM would cost, making it a highly cost-effective solution for local AI testing.
The physical build itself uses surprisingly modest components. Alongside the 768 gigabytes of Optane memory, the PC runs on an older Intel Xeon Gold 6246 processor and a Tyan motherboard. For memory caching, the system uses 192 gigabytes of standard Samsung DDR4 memory. For the graphics card, the builder did not use an expensive enterprise card. Instead, they plugged in a single Asus Dual GeForce RTX 3060, a budget-friendly gaming card with just 12 gigabytes of VRAM that you can easily buy online for around $300.
The builder configured the system’s memory settings to run in what Intel calls “Memory Mode.” In this mode, the operating system sees the massive 768 gigabytes of Optane persistent memory as the main system RAM. Meanwhile, the faster 192 gigabytes of standard DDR4 memory acts as a high-speed cache. The software then uses this massive memory pool to hold the bulk of the AI model’s data.
The model itself, Kimi K2.5, uses a modern “Mixture of Experts” architecture. To make the model fit on the cheap system, the builder used a highly compressed “quantized” version of the software. Using an open-source tool called llama.cpp, the builder strategically placed different parts of the AI on different hardware components. They loaded the core attention weights and routing components onto the cheap 12-gigabyte graphics card, while the remaining bulk of the model lived on the slower, Optane-backed system memory.
This hybrid setup worked beautifully. The system processed prompts at 16.44 tokens per second and generated new words at roughly 4 to 5.35 tokens per second. In a world where renting a single high-end Nvidia chip on the cloud can cost up to $40 an hour, running a trillion-parameter model locally on a home-built computer for a tiny fraction of that cost is a massive victory. It proves that clever memory management can bypass the hardware bottlenecks that hold back many developers.
This home project highlights a major shift in how the tech industry views AI hardware. For years, companies have focused almost entirely on buying faster graphics chips. However, researchers are realizing that memory capacity is actually the biggest bottleneck for running large models. If a system does not have enough memory to hold the model’s files, the graphics chip just sits idle. Using older, discontinued enterprise hardware like Optane could help small startups and researchers bypass expensive tech monopolies.
While Optane is no longer in production, the experiment shows that local AI innovation does not have to be limited to those with massive corporate budgets. It opens the door for a more democratic future where independent developers can run powerful tools at home. As more open-source models hit the web, finding creative ways to reuse older hardware will be key to keeping tech open and accessible. For now, APFrisco’s home server proves that a little creativity and some cheap, used parts can bring a trillion-parameter brain right to your desk.
