Paradigma open-sources a math-specialized model that achieves 94% on the AIME with only 1 billion parameters.

Beating AI News (from Dongcha) reports that Italian AI lab Paradigma has launched its first model, Limite 1B - Violetto. Boasting just 1 billion parameters, it is purpose-built for solving mathematical problems. Trained from scratch on fewer than 300 billion tokens, the model was further refined via supervised fine-tuning (SFT) and reinforcement learning. In official benchmark evaluations, Violetto scored 94.01% on AIME 2026, outperforming multiple models with parameter counts ranging from 3B to 120B. That said, some rival models’ results were sourced directly from their respective model cards or MathArena, rather than re-evaluated uniformly by Paradigma, so evaluation setups may differ. Paradigma positions Violetto as a high-throughput math model, prioritizing a small, specialized design. It primarily handles single-turn math queries and is also intended to act as a high-frequency math solver in multi-agent systems. The model weights and vLLM inference code are open-source under the Apache 2.0 license.