Paradigm Releases First Model Limite 1B, Targeting High-Frequency Mathematical Intelligence
On September 22, crypto firm Paradigm announced its first AI model, Limite 1B (codename Violetto), positioned as a "high-frequency mathematical intelligence" model focused on high-throughput mathematical reasoning. It is a 1B-parameter dense autoregressive transformer pretrained from scratch on under 300B curated tokens with a maximum context of 131k, and scored 74.25% on BeyondAIME—reportedly surpassing 30B-class models. The team has promised more variants to come.
Confirmed
- Limite 1B (Violetto) is Paradigm's first model: 1B parameters, dense autoregressive transformer, pretrained from scratch on fewer than 300B tokens, 131k max context
- Positioned for high-frequency mathematical intelligence, designed for single-turn responses, aiming at maximum per-parameter math capability
- Scored 74.25% on BeyondAIME, claimed to beat 30B-parameter models
- Post-training uses deliberately light instruction tuning, challenging the assumption that models must embed an assistant persona; users effectively get a near-base model
- The team teased more variants ("serving many more colors")
- The official account commissioned an artwork titled "limite" as the release visual, presenting the launch in an artistic manner
Why it matters
As Paradigm's first cross-domain release from crypto into AI, the model's "small model + minimal fine-tuning + per-parameter math capability" approach contrasts sharply with the mainstream trend of stacking parameters and heavy alignment. Commenters called it an impressive first release aimed at high-leverage use cases; researchers including YouJiacheng praised the sub-300B-token data efficiency as "amazing," sparking discussion on data efficiency of small models; others joked that its size is "perfect—big models are scary," while Dorialexander argued Europe is fully capable of training excellent small models that will keep improving. If independently verified, its BeyondAIME result offers a new reference point for lightweight specialized models.
2026-09-22 ~ 2026-09-22 · 12 related posts
Primary sources
- [source] Paradigm launches first model Limite 1B "Violetto" for high-frequency mathematical intelligence — tensorqt · 2026-09-22
- Paradigma's 1B model scores 74.25% on BeyondAIME, trained on under 300B tokens — tensorqt · 2026-09-22
- Paradigm's Limite model skips heavy instruction tuning to maximize math per parameter — tensorqt · 2026-09-22
- Paradigma's first model Limite 1B targets high-frequency math intelligence — tensorqt · 2026-09-22
- Paradigm launches Limite 1B Violetto, a tiny model for high-frequency math intelligence — tensorqt · 2026-09-22
- Paradigm Launches Limite 1B Math Model Trained From Scratch on 300B Tokens — tensorqt · 2026-09-22
- Paradigm unveils Limite 1B Violetto, a mysterious "high-frequency mathematical intelligence" model — tensorqt · 2026-09-22
- Paradigm launches Limite 1B, a tiny model built for high-frequency math reasoning — tensorqt · 2026-09-22
- Limite ships 1B math model trained on under 300B tokens — tensorqt · 2026-09-22