Kimi K3 CPU inference

Name: Kimi K3 CPU inference
Uploaded: 2026-08-08T14:32:01.000Z
Duration: 20 s
Channel: Vojtech
Description: Kimi K3’s 2.78 trillion parameters just got squeezed onto a single CPU with 8.24GB of RAM. The engine is 176KB of pure C99, no GPU, no CUDA, no framework in sight. It’s a mixture-of-experts setup with 896 experts, but only 16 fire per token. The rest stream from an NVMe drive on demand, so 93% of the model never touches RAM. Open source, Apache 2.0, repo is kimi-k3-in-c.

by Vojtech: Kimi K3

Kimi K3’s 2.78 trillion parameters just got squeezed onto a single CPU with 8.24GB of RAM. The engine is 176KB of pure C99, no GPU, no CUDA, no framework in sight. It’s a mixture-of-experts setup with 896 experts, but only 16 fire per token. The rest stream from an NVMe drive on demand, so 93% of the model never touches RAM. Open source, Apache 2.0, repo is kimi-k3-in-c.

Transcript (en)

God, you look so pretty when you tell me that you love me. I wish that I could lie, but my mind gets in the way. I know you think that I'm always way too self-aware. Oh, it could never be too...

Kimi K3 CPU inference

Transcript (en)

God, you look so pretty when you tell me that you love me. I wish that I could lie, but my mind gets in the way. I know you think that I'm always way too self-aware. Oh, it could never be too...