GitHub radar
Kimi K3 in C: 2.78T-Param LLM on CPU using 8 GB of RAM
Runs Moonshot AI’s Kimi K3 (2.78T parameter mixture-of-experts language model) on a single CPU using only 8.24GB of RAM while streaming a 1.56TB checkpoint from disk. No GPUs, frameworks, just 176KB of C99 code.
C99 implementation of Moonshot AI’s Kimi K3 (2.78T parameter mixture-of-experts language model) running entirely on CPU without BLAS or CUDA or any ML framework. Streaming a 1.56TB checkpoint from NVMe while keeping only 8GB of active experts in memory. Achieves identical result across different memory sizes (from 8GB to 224GB).
Why a vibe-coder should care
Demonstrates how we can run state-of-the-art AI models using commodity hardware (such as your laptop) rather than specialized GPUs. Shows that by using smart streaming + quantization methods, we can all do trillion parameter inferences at home!
▌ More finds