DatarusAI Launches 14B Reasoning Model That Thinks like an Analyst–Sketching, Revising, & Solving Complex Problems with Human-Like Efficiency
Paris, France – August 21, 2025 – DatarusAI today announced the release of Datarus-R1, a groundbreaking 14-billion parameter reasoning model that delivers the performance of 32-billion parameter systems while using 18-49% fewer computational tokens, potentially transforming the economics of enterprise AI deployment.
Currently trending 15th on Hugging Face’s global platform, Datarus-R1 addresses a critical industry challenge where modern reasoning models generate over 20,000 tokens of circular reasoning for tasks requiring far less computation. The Paris-based team’s breakthrough demonstrates that European AI innovation can compete directly with Silicon Valley giants by focusing on efficiency rather than scale.
The team trained Datarus-R1 on 144,000 synthetic analytical trajectories across finance, medicine, and numerical analysis, combined with curated reasoning datasets specifically calibrated to prevent overthinking patterns. This approach included 10% complete failures in the dataset, teaching the model to recognize and avoid unproductive reasoning paths. The result is an “AHA-moment” pattern where the model forms an initial hypothesis, identifies issues, revises once or twice, then converges on the solution without excessive contemplation.
Performance benchmarks reveal remarkable achievements despite the model’s smaller size. Datarus-R1 scores 57.7% on LiveCodeBench v6. It achieves 70.1% on AIME 2024, the highest score among 14B-parameter models, and 62.1% on GPQA Diamond, demonstrating graduate-level scientific reasoning.
On complex problems where competitors balloon from 2,000 to over 20,000 tokens, a 945% increase, Datarus-R1 maintains stable token usage around 5,400 tokens. This efficiency translates to potential cost savings of 80-90% for enterprise deployments, making previously uneconomical automation suddenly viable at scale.
Datarus-R1 operates as a virtual data analyst capable of automatically loading and examining data structures, generating statistical hypotheses, writing and executing analysis code, handling errors, and producing visualizations and reports. The system offers two operational modes: Agentic Mode for interactive analysis with live code execution, and Reflection Mode for generating concise documentation, suitable for both automated pipelines and human-readable reporting.
In line with European commitment to open AI development, DatarusAI has released the model under an Apache 2.0 license, making it freely available for research and commercial use. The release includes full model weights on Hugging Face, complete training methodology, and the Datarus-JupyterAgent repository on GitHub that generates complete analysis notebooks from a single prompt. This one-command deployment makes sophisticated data analysis accessible to teams without extensive ML engineering resources.
For enterprises facing spiraling AI inference costs, Datarus-R1 represents a paradigm shift. Tasks previously too expensive to automate, including routine data analysis, report generation, and statistical validation, become viable at scale. By proving that a small team can build something competing with models from organizations hundreds of times their size, Datarus has opened new possibilities for AI innovation outside traditional tech giants.
The complete model, code, and research paper are available immediately, with the team actively seeking developer community feedback.
About DatarusAI
DatarusAI is a Paris-based artificial intelligence research team dedicated to developing efficient, accessible AI solutions that solve real-world problems with minimal computational overhead. The company specializes in creating models that deliver maximum performance while dramatically reducing operational costs.
Technical Specifications
Parameters: 14.8 billion
Training: 144,000 synthetic analytical trajectories plus curated reasoning datasets Token Efficiency: 18-49% reduction versus competitors
License: Apache 2.0 open source
Availability: HuggingFace, GitHub Datarus-JupyterAgent, Demo at chat.datarus.ai
Media Contact
Contact Person Name: Ayoub Ben Chaliah
Contact Email: ayoub.ben-chaliah@clairechains.com
City & Country: Paris, France
