kimi-k3-in-c

Why we included this project

Running a 2.78-trillion-parameter model is normally a job for a cluster, but this project does it on a single CPU with about 8 GB of RAM. It is a from-scratch inference engine in portable C99, with no BLAS, no deep-learning framework, and no GPU. The trick is that it streams most of the 1.56 TB checkpoint off disk on every step and keeps only a selected trunk resident in memory, so output stays byte-identical no matter how much memory you give it; the budget only changes the clock. The cost is speed and convenience: this is the base model, so you get continuations rather than chat replies, and each token takes seconds instead of milliseconds. For systems programmers curious about memory-efficient MoE inference, packed 4-bit expert routing, or how far a plain CPU can be pushed, that trade-off is the point.