Bonsai LM (1-bit and 1.58-bitLLMs) benchmark on Jetson Orin Nano Super Just released a deep benchmark of 5 Bonsai LM models (1.7B → ~8B) on a $250 Jetson Orin Nano Super 8GB using llama.cpp CUDA - across all 4 power modes: 7W, 15W, 25W, and MAXN A thread! So, Bonsai LM models are new line of 1-bit LLMs released recently and I was wondering how they perform in terms of TTFT, tok/s, tok/J and overall request latency, with incredibly low memory footprint even for 8B models! Thus, I ran a ...