Tools
A 176 KB C99 Engine Runs 2.78T-Parameter Kimi K3 in 8.24 GB of RAM With No BLAS, Framework, or GPU
kimi-k3-in-c (FareedKhan-dev, 691 stars, created 2026-08-01, Apache-2.0) reports 8.24 GB measured peak RSS running inference against a 1.56 TB checkpoint, with the whole engine at 176 KB of portable C99 targeting Linux x86-64. The tradeoff is stark and honestly reported: 32.69 seconds per token averaged over 8 tokens, and it is the base model with no chat template, so prompts get continuations rather than replies. Alongside WASTE and Deltafin, it is the third independent take in a week on the same question of how far frontier-scale MoE can be pushed onto one machine.
Source
↳ Follow the thread