Kimi K3 vs Abacus Supercomputer: local coding comparison

The speaker gives Kimi K3 and Abacus AI Supercomputer the same prompt for a TypeScript front end called Silicon Exchange. It uses mock data and asks for dark mode, accessibility, microinteractions and unit tests. Abacus sponsors the video.

The local LLM runs across four Mac Studios with 512 GB of unified memory each, connected through a Thunderbolt 5 mesh with RDMA. The speaker uses MLX distributed and points OpenCode at the cluster. He describes an unpruned quantized model with all 896 experts and 817 GB of weights, with extra memory needed for context. He reports prompt processing at about 238 tokens per second and generation at 14.7 tokens per second on this setup.

The initial local attempt produces no visible tokens while he waits. A smaller request returns a text box and button app in about three minutes. After retrying the full prompt and leaving it running, he reports a completed interface four hours later. He gives Abacus's completion time as 15 minutes. These are results from his demonstration, rather than a general benchmark.

Supercomputer supplies an always-on cloud VM with terminal, desktop and file access. The cloud run selects Opus 5 high and GPT 5.6 Soul, so the comparison covers different models and execution environments. A separate Abacus AI Desktop example uses hosted Kimi K3 to write a local Node application. The speaker favors the cloud coding assistant for quick tasks and the cluster when data must stay in the building.