User Reports Impressive Performance for Liquid's LFM3-1.7B Model on AMD Hardware

A user shared real-world performance metrics for Liquid AI's LFM3-1.7B model running via llama.cpp and ROCm on AMD hardware. The model demonstrated strong coherence and high processing speeds across large context windows. This report highlights the viability of running small, efficient language models on edge and consumer AMD hardware for tasks like text compression and summarization. It showcases how recent model generations achieve surprising coherence despite their compact size. Tested using the Q8_0 GGUF format, the model achieved about 4,100 tokens per second prefill and 78 tokens per second decode on an AMD Strix Halo iGPU. Tensor-splitting the workload across a Radeon 7900 XTX and 7900 XT further boosted performance to roughly 7,100 prefill and 151 decode tokens per second while maintaining stability at 32k to 64k token prompts.