Paradigma published its `limite-1b-violetto` checkpoint and tokenizer on Hugging Face, making a compact model for mathematical reasoning publicly available.
The Hugging Face repository identifies Violetto as a dense 1-billion-parameter model with support for contexts up to 131k tokens. Paradigma says it was trained from scratch on fewer than 300 billion curated tokens and reports a 74.25% average on BeyondAIME, while cautioning that evaluation setups may differ.
The weights and accompanying vLLM serving code are licensed under Apache-2.0. Builders can self-host the model through Paradigma’s custom vLLM plugin, though direct Transformers support is still pending.
Violetto is intended for single-turn math problems rather than general assistant use. Paradigma also warns that it may reinterpret prompts instead of reliably following broader instructions.