Paradigma open-source 1000000000-parameter mathematical model Violetto, officially reported to have AIME 2026 scores 94.01%
BIBIBI
AT A GLANCE
Paradigma released Violetto, a -parameter mathematical model, claiming that it scored AIME 2026 points 94.01%。
Article
-parameter AIME 94%, Paradigma open-source math-specialized model
Beating AI news, Italian AI laboratory Paradigma has released its first model, Limite - Violetto. It has only 1000000000 parameters and is designed specifically for solving math problems. The model was trained from scratch using fewer than 300000000000 tokens tokens, followed by SFT and reinforcement-learning post-training.
According to the benchmark results published officially, Violetto achieved AIME 2026 achieved 94.01%on vLLM and its Apache 2.0 inference code have been released under the https://m.theblockbeats.info/flash/368445
Key points
01
Paradigma releases its first model, Limite – Violetto.
02
Violetto has 1000000000 parameters and is designed specifically for solving math problems.
03
The model was trained from scratch, using fewer than 300000000000 tokens, then underwent SFT and reinforcement-learning post-training.
04
The official evaluation table shows that Violetto scored AIME 2026 points 94.01%。
05
Violetto is positioned as a high-throughput mathematics model, primarily handling single-turn mathematics problems.
06
Violetto is planned to serve as a high-frequency mathematics solver in a multi-agent system.
07
The model weights and vLLM inference code have been released under the Apache 2.0 license.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This is a relatively small-parameter model specialized in solving mathematics problems. According to the evaluation results reported, it AIME 2026 achieved 94.01%.
Why it matters to readers
It demonstrates an approach in which small-parameter models focus on a single task, and provides open weights and inference code.
It can be understood as a smaller open-source model specialized in mathematics problems.
Risks and unknowns
Some competitors' results were taken directly from model cards or MathArena, rather than being retested uniformly by Paradigma.
Evaluation configurations may differ across models, so direct comparisons between results are subject to uncertainty.
Related Developments
Loading event timeline…
Related concepts
Parameters
This term is not in the glossary yet. Browse related concepts in the glossary.
