Paradigma Launches 1B-Parameter Math-Specific Model Violetto, Achieves 94.01% on AIME 2026

ME News report, September 22 (UTC+8), Beating AI Bulletin: Italian AI lab Paradigma has released its first model, Limite 1B - Violetto. With only 1 billion parameters, it is specifically designed to solve mathematical problems. The model was trained from scratch using fewer than 300 billion tokens, followed by SFT and reinforcement learning. According to official evaluation results, Violetto achieved 94.01% on AIME 2026, outperforming multiple models ranging from 3B to 120B parameters. However, some competing scores were directly taken from model cards or MathArena and were not re-evaluated uniformly by Paradigma, meaning evaluation configurations may differ. Paradigma positions Violetto as a high-throughput mathematical model, focusing on making the model smaller and more specialized. It primarily handles single-turn math problems and is also planned to serve as a high-frequency math solver within multi-agent systems. The model weights and vLLM inference code have been open-sourced under the Apache 2.0 license. (Source: BlockBeats)