The repository describes the target as a 105 GB AI model on Macs with 16 GB to 64 GB of memory. It is written in native Swift and uses MLX and Metal; the project says it can run offline, without Python. The repository also lists a CLI and a Swift library.
The source frames the approach around mixture-of-experts inference, where experts can be loaded from SSD as needed. It does not establish independent throughput, latency, quality, SSD-wear, stability, or results across the stated Mac memory configurations. It also does not, in the supplied material, provide a comparative benchmark against other local runtimes.