Limite 1B - Violetto

Not yet independently verified. The evaluation results are Paradigma’s own and have not been reproduced independently; some comparison figures in its table were taken from other model cards or MathArena rather than re-run. We will update this when it can be confirmed, and remove this note.

Limite 1B - Violetto is the first model from Paradigma, a research lab whose release credits eight authors, among them Emanuele Rodolà and Giulio Starace. It is a small language model built for high-throughput solving of difficult mathematics problems and is meant to answer single-turn prompts.

Paradigma released it on 21 September 2026: its announcement on X, the Hugging Face repositories for the model, its base checkpoints and a value model, and the GitHub repository all date from that day, and RuntimeWire reported the release the same day. The blog post’s metadata carries 18 September, which no other source supports.

Limite is a 1-billion-parameter dense autoregressive transformer trained from scratch on fewer than 300 billion curated tokens, and handles sequences of up to about 131,000 tokens. Paradigma used synthetic data, supervised fine-tuning and reinforcement learning, and took architecture ideas from community pre-training speed-run competitions. It is served through a custom vLLM plugin on Linux with an NVIDIA GPU; Transformers support is described as coming soon. The weights and serving code are released under Apache 2.0, together with the value model used in training.

In its own evaluation across seven mathematics benchmarks, Paradigma reports an average of 74.25% on BeyondAIME; some comparison figures in its table were taken from model cards or MathArena rather than re-run. A technical report is promised.

The model is deliberately lightly instruction-tuned. Paradigma's own examples show it misreading ordinary questions, such as what photosynthesis is, as mathematical tasks, so it is not suited to general assistant use or instruction following. It is best treated as a component for mathematical reasoning pipelines.