LFM3-1.7b seems impressive

Been testing Liquid's LFM3-1.7B (Q8_0 GGUF, llama.cpp, ROCm) as a candidate for compression and summarization work 4.1k t/s prefill and ~78 t/s decode on a Strix Halo iGPU ~7.1k / ~151 if you tensor-split it across a 7900 XTX + XT its way more coherent at decent sized prompts (32k-64k) then 1.7b deserves to be imo