Paris, France — August 22, 2025 — Leads & Copy — DatarusAI has announced the release of Datarus-R1, a 14-billion parameter reasoning model that matches the performance of 32-billion parameter systems while using 18-49% fewer computational tokens. Trending 15th on Hugging Face, Datarus-R1 addresses challenges related to excessive computational tokens.
The model was trained on 144,000 synthetic analytical trajectories across finance, medicine, and numerical analysis, teaching it to recognize and avoid unproductive reasoning paths. Datarus-R1 scores 57.7% on LiveCodeBench v6, 70.1% on AIME 2024, and 62.1% on GPQA Diamond.
Datarus-R1 maintains stable token usage around 5,400 tokens on complex problems, translating to potential cost savings of 80-90% for enterprise deployments. It operates as a virtual data analyst, offering Agentic Mode for interactive analysis and Reflection Mode for documentation.
DatarusAI has released the model under an Apache 2.0 license, including full model weights on Hugging Face, complete training methodology, and the Datarus-JupyterAgent repository on GitHub. The complete model, code, and research paper are available immediately.
DatarusAI specializes in creating efficient, accessible AI solutions. The company is based in Paris.
Ayoub Ben Chaliah can be reached at ayoub.ben-chaliah@clairechains.com
Source: DatarusAI
