Kimi k3 now runs on one consumer GPU

Kimi now runs on one consumer GPU Kimi-Linearis a 48B model that normally needs a datacenter card. we shrank it to 28.8 GB and made it decode at 113.83 tok/s, so it fits one 32 GB GPU. an RTX 5090 runs it it works with the coding agent you already use. Claude Code, Codex, Cline, Aider, opencode. point it at the local server and it just works, because k3 detects which client is calling and translates for it nothing leaves your machine. no API key, no per-token bill, no rate limit. Apache-2.0 24 GB cards are next!! the technical part: two W4A16 kernels held 77% of decode at 8 to 10% of the card's memory bandwidth. both were written as matrix-matrix products and used at decode as matrix-vector products. rewriting them as real GEMVs is the entire 3.18x. every number names the hardware it ran on and ships the runner that produced it, including the three fusions that made things slower