AnalysisAI ModelsSeptember 15, 2026

Developer trains 44M-parameter quantized LLM that runs at ~1,900 tok/s on CPU

The model ships in 19.8 MB after training on 45B tokens. It follows the developer's earlier SHADOW-250M, a 60 MB model that ran ~400 tok/s on CPU but could not reliably reason over retrieved records or compute.

1 source

More stories today

Open the live feed