BitNet
Trained a 100M-Param Model for €100 — Training Got Cheap, Capabilities Didn't
Reddit user trained a 100M-param BitNet for €100 on 4 B300 GPUs. 40.59% on tests — beats GPT-2, trails major models. News is cost, not capability.
Back to home
BitNet
2 articles tagged with this topic
BitNet
BitNet hits 36 tokens/sec on a plain CPU — LLM inference starts shedding its GPU dependency
A developer shifu_legend wrote a zero-dependency inference engine in pure C99, hitting 36 tokens/sec on an Intel Xeon running a 1.58-bit BitNet model.