👀 A brand-new Local AI runtime is already beating llama.cpp on an 8GB RTX 3050 at least in some tests. SonicBoom v0.0.3 is a new compiler-first LLM inference engine from HardenedLinux. No Python at build time & no Python at runtime. And on a RTX 3050 8GB, its developers compared it against llama.cpp using the same GGUFs and same session setup. With Gemma 4 E4B Q3_K_M: ⚡ decode → 1.18× llama.cpp 📥 prefill → about the same With Qwen2.5-0.5B Q4_K_M: ⚡ decode → 1.12× llama.cpp 🚀 prefill → 1.09× It also does prefix reuse + execution-plan caching, which the developers say cuts TTFT by several times in their multi-turn agent workload. 🎯 Someone built a new inference engine and is already getting competitive performance on an 8GB consumer GPU. ⚠️ These numbers currently come from the developers themselves, on one RTX 3050. Now they're asking Localmaxxers with other GPUs to reproduce the benchmarks. Links in ALT