125B AI on a Gaming PC? Strata Tested on RTX 5090 Desktop & Laptop

Published 06/10/2026

Description

Can you run Qwen3.8-Flash-Next locally without a server-class GPU? In this hands-on Strata walkthrough, I install the inference engine on Windows, configure model quantization and context settings, and test local C# coding with an RTX 5090 desktop and an RTX 5090 laptop. Contents: 0:00 - Introduction 0:20 - Description 2:00 - Git 4:24 - Install 15:45 - UI 19:41 - UI Test 23:40 - Claude 32:25 - Update 33:45 - Conclusion Strata is designed to run Qwen3.8-Flash-Next on consumer hardware using both GPU memory and system RAM. In this video, I explore the standard Flash Next model, the Coder variant, and Swift 1.5, then compare simple web-chat responses with complete Claude Code workflows. The walkthrough covers: Installing Strata and downloading the engine and model files Choosing Q2, IQ2, and IQ3 quantizations Configuring context length, KV-cache precision, and optional image support Running the automatic hardware-tuning process Editing startup scripts, ports, host binding, and API-key settings Using the web interface to monitor tokens per second, VRAM, RAM, GPU temperature, and power Testing image understanding and exporting responses as Markdown Connecting Claude Code and checking whether generated .NET applications build and run Updating Strata as new releases become available The coding-agent results varied considerably in my tests. The Coder variant failed to complete the task successfully, while the standard Q2 run took approximately 56 minutes. Swift 1.5 finished in around 5 minutes and produced a working application, but did not follow my one-file-per-class rule. The Flash Next IQ3_S run completed in 4 minutes 5 seconds and produced a working application with separate project folders. On my RTX 5090 laptop, IQ2_XS completed the task in 6 minutes 26 seconds. These are results from my hardware, prompt, settings, and the Strata versions tested—not a universal benchmark. High token-generation speed is useful, but completing