Limite 1B Violetto: A 1-Billion-Parameter Math Solver

Italian startup Paradigma has released Limite 1B Violetto, a dense autoregressive transformer with 1 billion parameters, trained from scratch on fewer than 300 billion curated tokens and achieving 94.01% on AIME 2026 with a compute budget 1–3 orders of magnitude lower than models of the same level.

Compiled by Sergey KostenchukPublished 2026-09-23Updated 2026-09-23

2026-09-23 Research NVIDIA

🤖 Limite 1B Violetto: A 1-Billion-Parameter Math Solver

Italian startup Paradigma trained a dense 1-billion-parameter transformer from scratch: fewer than 300 billion curated tokens in 6 weeks, context up to 131,072 tokens. Not a chatbot: the input is a problem, the output is a solution. On AIME 2026 — 94.01%, on BeyondAIME — 74.25%.

🌍 The model is tens of times smaller than competitors at the same level, with 1–3 orders of magnitude less compute: an argument in favor of speedrun pretraining and data curation instead of scale. The “solver + Limite Value Model” combination is a block for best-of-n and RL.

👤 Weights are open on Hugging Face (Apache 2.0): a local olympiad math solver — ~2 GB of weights, ~7.5 GB VRAM. Runs via a vLLM plugin on Linux with NVIDIA. The prompt must be a math problem: ordinary questions are reinterpreted as mathematical.

Source 1: https://paradigma.inc/blog/limite-1b-violetto/ Source 2: https://huggingface.co/paradigma-inc/limite-1b-violetto

Sources