Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s ---- - 28 minutes ago | 18 points | 4 comments - URL: https://github.com/Niko1221/Strata - Discussions: https://news.ycombinator.com/item?id=49953495 - Summary: Strata is an open-source inference engine that runs the 125-billion-parameter Qwen3.8-Flash-Next AI model on consumer PCs with NVIDIA or AMD GPUs (12GB+ VRAM), 32GB+ RAM, and 80GB disk. It offers one-click install for Windows/Linux, with setup via AI assistants or manual scripts. The engine distributes work across GPU, RAM, CPU, and SSD, achieving 44–94 tokens/s generation and 1,100–2,650 tokens/s prompt processing on tested gaming PCs. Users can select model sizes (e.g., Q2_0, IQ2_XS, IQ3_S, Coder) based on RAM capacity. It provides a local web app at localhost:8080, OpenAI/Anthropic-compatible APIs, optional image input, and multi-GPU support. Troubleshooting covers common issues like freezes, download interruptions, and port conflicts. The project is MIT-licensed, credits Qwen, ISTA-DASLab, and others, and includes detailed documentation.