1B model scores 94% on AIME 2026: Paradigma open-sources Limite 1B Violetto
tensorqt · x · 2026-10-07
Paradigma released the technical report and open-sourced (Apache 2.0) its first model, Limite 1B - Violetto, built to maximize mathematical reasoning per parameter and per joule.
- Scale & data: 1B parameters trained from scratch on a curated math-centric corpus of 237B tokens, with up to 131k context; 6 weeks from first experiments to release.
- Recipe: post-training combines distillation, targeted SFT and RLVR on a curriculum; pre-training interleaves SFT with DPPO RL, with architecture borrowed from nanogpt speedrun (NorMuon, MUDD variant, XSA).
- Results: 94.01% on AIME 2026 and 74.25% on BeyondAIME, matching models many times its size on their eval suite.
- Positioning: deliberately lightly instruction-tuned, meant as a single-turn math solver and a fast parallel component in larger systems rather than a general assistant.
- Model, base and value checkpoints, and an optimized serving implementation are released.
Related event: Paradigm Open-Sources Limite 1B - Violetto, Its First Math-Focused Model(6 posts)→
More from Models
- Gemini Ultra's $100 plan includes 1000 monthly Colab credits—about 55 H100 hours — iskander · 2026-11-22
- LeCun: Training LLMs on human knowledge yields memory, not intelligence — haider1 · 2026-11-22
- Claude veteran shocked by Gemini's leap, weighs switching to Google AI Ultra — ScreenPlayLife · 2026-11-22
- Gemini Neon builds an interactive USL lock-contention UI exposing a testing blind spot — tianyin_xu · 2026-11-22
- New spatial planning benchmark: GPT-6 and Claude Opus 5.5 fail parking route tasks — songhan_mit · 2026-11-22
- Microsoft's new models reportedly up to 80x cheaper and 50x faster — ChrisUniverse · 2026-11-22