Strata Runs 100B-Parameter Models on 8GB GPUs

On October 8, 2026, an author named Wu Jiahao published a technical article describing a scheme called Strata that runs a 125B-parameter MoE large model on a single RTX 5090 with 32GB of VRAM, tested with driver version 617. The article framed the approach as bringing server-class MoE inference to ordinary PCs. The post appeared on Juejin on October 9 at 22:51.

Early the next day, on October 10 at 01:12, the same author published a follow-up tutorial on Juejin covering the full workflow for running the 125B model on a single RTX 5090 with 32GB of VRAM. It walked through installation, verification, and benchmarking, and listed a more specific test environment that included driver version 617.14.

Later that day, at 16:44 on October 10, the site Xiaozhong Software reported that it had tested Strata itself and found the tool can run a trillion-parameter-class large model on an ordinary computer with only 8GB of VRAM. The report said the approach was verified on a range of GPUs, including the RTX 5070, RTX 3060, RTX 4060 Ti, RX 9070 XT, RX 9060 XT, RX 7800 XT, RTX 3090, and even older Quadro cards. According to that report, this significantly lowers the GPU barrier for running very large models locally. The public information currently stops there: the reports establish that Strata is feasible, but they do not name the developer behind the scheme, describe its implementation in detail, or mention any next steps.