Fractal Unveils Fathom-R1-14B Reasoning Model on DeepSeek

Fractal has introduced Fathom-R1-14B, a new open-source large language model, available for $499. The model, which boasts 14 billion parameters, is designed to deliver superior mathematical reasoning performance, surpassing models like o1-mini and o3-mini, and approaching the capabilities of o4-mini. The model can be accessed on Hugging Face, with its codebase available on GitHub under the MIT license.

Fathom-R1-14B was developed as part of an initiative to create India's first large reasoning model under the IndiaAI mission. It achieves notable performance on olympiad-level exams, with Pass@1 accuracy scores of 52.71% on AIME-25 and 35.26% on HMMT-25. These scores improve significantly with additional inference-time compute.

The model was post-trained using supervised fine-tuning, curriculum learning, and model merging. Fractal has also released a variant, Fathom-R1-14B-RS, which uses reinforcement learning and supervised fine-tuning, available at a higher cost of $967.