Paradigma's 1B model scores 74.25% on BeyondAIME, trained on under 300B tokens
tensorqt · x · 2026-09-22
Paradigma released Limite 1B (Violetto), a 1B-parameter dense autoregressive transformer trained from scratch on under 300B curated tokens with 131k context. Using synthetic data, curated SFT and RL post-training — plus an architecture inspired by pre-training speedrun advances — it averages 74.25% on BeyondAIME, beating the 30B MUSE-Glimmer at 70%. Deliberately lightly instruction-tuned for single-turn math use, it ships with model weights, a training value model, and a custom vLLM inference plugin; a tech report is coming.
More from Models
- Gemini beats GPT-6 Astra at robot capture the flag, winning 70% of matches — chris_j_paxton · 2026-09-22
- Jev ships completions API — but its maker says don't use it for text generation, only scoring — AlexandrePesant · 2026-09-22
- Indie dev says he built non-autoregressive decision models a year before TypeSafe AI's Jev — HowDevelop · 2026-09-22
- 105 Planted Bugs Put Grok 4.7 at 28.7 vs GPT-6 Astra's 45 in Real-Repo Coding Test — PawelHuryn · 2026-09-22
- Regression on OpenAI's GPT-5.6 Benchmark Data Backs Out Their λ Values — tobyordoxford · 2026-09-22
- Paradigm Launches Limite 1B Math Model Trained From Scratch on 300B Tokens — tensorqt · 2026-09-22