Skip to content
D-CSIL

Research Domain

Local Intelligence

Foundation models trained end-to-end on hardware the lab owns — M3 Max unified memory, custom Metal kernels, memory-mapped corpora — plus edge TTS on Raspberry Pi.

MEASUREDCURRENT EVIDENCE LEVEL

Research Question

How much frontier-style capability can be built with zero rented compute?

The D-CSIL-3 pipeline pretrains a 235M selective-state model on a 4.75B-token, seven-stage all-local corpus (FineWeb-Edu, Wikipedia, SlimPajama, OpenWebText, WikiText-103, SFT mix) at a measured 1,270 tokens/sec.

The mlx-recurrence kernel family exists because local training demanded it: fused Metal scans deliver up to 31.8× forward+backward speedups and were hot-swapped into a live run, cutting peak memory from 23.9GB to 10.3GB.

Experiments in this domain