Gemma-4-E2B writes 38 tok/s vs 19 for llama.cpp, 26B-A4B 18 vs 12, on a 32 GB laptop. Gemma 4 models converted for Janas-LLM.
AI & ML interests
Janas-LLM: 37 tok/s from Qwen3-Next-80B on a 32 GB laptop (24 without drafts; llama.cpp on the same laptop: 2.8). Pure-C inference for LLMs bigger than your RAM, CPU or GPU.