Limite 1B - Violetto: a one-billion-parameter model aimed at high-throughput math reasoning
The technical report describes a small model trained from scratch on a mathematics-focused corpus, with checkpoints and serving code released under Apache 2.0.
In plain words
A research team says it built a very small AI model specialized for solving math problems quickly and cheaply. It was trained only on math-heavy text, and the authors report strong benchmark scores, though the model is not meant for general conversation.
What was announced
According to the technical report on arXiv, Limite 1B - Violetto is a one-billion-parameter model built for high-throughput mathematical reasoning. The authors state it was trained from scratch on 237 billion tokens drawn from a curated, mathematics-centric corpus and handles context lengths up to 131k tokens. Architecture and training recipe were reportedly designed together to maximize reasoning while keeping training and inference costs low. Post-training combined knowledge distillation, targeted supervised fine-tuning, and RLVR on a curated curriculum.
On the authors' own evaluation suite, the model reportedly scored 94.01% on AIME 2026 and 74.25% on BeyondAIME, which they say places it among models many times its size. The report stresses that the model is only lightly instruction-tuned: it is intended as a single-turn math solver or a fast parallel component inside larger systems, not a general-purpose assistant. The team released the model, base and value checkpoints, and an optimized serving implementation under the Apache 2.0 license.
What it does not show
- All figures and benchmark results come from the authors' own abstract; no independent evaluation is available in the supplied source.
- The source is an abstract only, so details of the architecture, training data curation, and evaluation methodology are not described.
- The comparison with much larger models reflects the authors' evaluation suite and is not independently confirmed.
For security and IT teams
- Scope
- The released artifacts cover the model, its base and value checkpoints, and an optimized serving implementation, licensed under Apache 2.0; the model is positioned as a single-turn math solver rather than a general assistant.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.