Custom 250M LLM: 60MB Deploy, 100M-Token Disk Context Retrieval

Final-Data-1410 · reddit · 2026-08-22

Trained a 250M parameter model from scratch on 30B tokens, quantized to <2 bits for a 60MB deployment. It runs at 400 tok/s on CPU. Its key feature is a long context mechanism where tokens older than 2048 are compressed to 1 bit and stored on disk (320 bytes/token), supporting retrieval from up to 100M tokens of history. The vocabulary uses fixed 512-bit codes with zero trained parameters. The model successfully retrieved a serial number from 50.6M tokens deep. Code and weights are open-sourced.

Original post →

More from coding & agent

coding & agent channel →