Morning Edition · Thursday, September 24, 2026Published at 2:18 AM EDT · New York

An Italian Lab's 1-Billion-Parameter Model Reports 94 Percent on AIME 2026, Trained on Under 300 Billion Tokens

Paradigma released Limite 1B Violetto under Apache 2.0 with a 131,072-token context, a dense transformer built for one domain and nothing else.

Paradigma, an Italian artificial intelligence laboratory, published Limite 1B Violetto, a dense autoregressive transformer (a type of neural network architecture) of roughly one billion parameters with a context window of 131,072 tokens, tr…