Community · Nov 1, 2026, 15:38 UTC

Liquid's LFM3-1.7B shows strong prefill and decode speeds on a Strix Halo iGPU

A Reddit user reports about 4.1k tokens/s prefill and 78 tokens/s decode for the Q8_0 GGUF on llama.cpp with ROCm, rising to about 7.1k and 151 tokens/s when split across two AMD GPUs. The user says output stays more coherent at 32k to 64k prompts.

• LabLiquid
• ModelLFM3-1.7B
• Topicopen-weights
• Topicbenchmark