Paradigma announced on 18 November 2026 that it is releasing Limite 1B – Violetto, a new language model built specifically for mathematical reasoning. The model has one billion parameters and is trained on 237 billion tokens drawn from a curated mathematics‑focused corpus. Unlike many large‑scale models that aim to be general assistants, Violetto is lightly instruction‑tuned and meant to be used as a single‑turn solver or as a fast component inside larger systems.
Violetto can process up to 131 k tokens in a single prompt, a length far beyond typical chat models that stop around 4–8 k tokens. This long context window lets it keep large problem statements, proofs, or code snippets in view while it works. The developers also released an optimized serving implementation, which they say keeps training and inference costs low compared with other math‑oriented models.
In benchmark tests the model scored 94.01 % on the 2026 American Invitational Mathematics Examination (AIME) and 74.25 % on BeyondAIME, a harder set designed to push reasoning limits. Those numbers put Violetto among models many times its size on Paradigma’s internal evaluation suite. The results were obtained after a post‑training pipeline that combined knowledge distillation, targeted supervised fine‑tuning, and reinforcement learning with human feedback (RLHF) on a curated curriculum.
Why release a small, math‑focused model?
The research team explains that a smaller model reduces hardware requirements, making it easier for universities or hobbyists to run math‑heavy tasks without needing a data‑center‑grade GPU cluster. Because the model is released under the Apache 2.0 license, anyone can download the checkpoints, adapt the architecture, or embed the serving code in their own applications. Paradigma hopes this will encourage more experimentation in automated theorem proving, homework help tools, and high‑throughput symbolic computation pipelines.
Why it matters
For students and educators, Violetto offers a free tool that can solve complex math problems without needing expensive cloud credits. Researchers can integrate the fast, parallel solver into larger AI pipelines, potentially accelerating work in automated proof generation. However, the model is only lightly tuned for dialogue, so it may not perform well as a general‑purpose chatbot.