Fractal, Mumbai primarily based AI firm, has launched a brand new open-source giant language mannequin, Fathom-R1-14B. The mannequin delivers mathematical reasoning efficiency that surpasses o1-mini and o3-mini, and approaches o4-mini ranges, all at a post-training price of simply $499.
The mannequin is on the market to strive on Hugging Face, and the codebase is on GitHub. It’s obtainable below the MIT license, together with datasets and coaching recipes.
Developed as a part of a proposed initiative to construct India’s first giant reasoning mannequin below the IndiaAI mission, Fathom-R1-14B is a 14-billion-parameter mannequin derived from Deepseek-R1-Distilled-Qwen-14B.
“We proposed constructing India’s first giant reasoning mannequin as a part of the IndiaAI mission. We proposed constructing three fashions (a small one, a mid-sized one and a big one with 70 billion parameters),” mentioned Fractal CEO Srikanth Velamakanni in a LinkedIn publish.
He additional added that “That is only a tiny proof of what’s doable.”
On olympiad-level exams AIME-25 and HMMT-25, Fathom-R1-14B achieves 52.71% and 35.26% Go@1 accuracy, respectively. When allowed further inference-time compute (cons@64), the scores rise to 76.7% and 56.7%.
“It delivers efficiency rivalling closed-source o4-mini (low) with respect to cons@64 ,all whereas staying inside a 16K context window,” the corporate mentioned.
The mannequin was post-trained utilizing supervised fine-tuning (SFT), curriculum studying, and mannequin merging.
“We carry out supervised fine-tuning on rigorously curated datasets utilizing a selected coaching method, adopted by mannequin merging,” the corporate mentioned.
Fractal has additionally launched a separate variant, Fathom-R1-14B-RS, achieved comparable outcomes utilizing a mixture of reinforcement studying and SFT, costing $967.
Final yr, the corporate launched Vaidya.ai, a multi-modal AI platform designed to supply free and accessible healthcare help. In the meantime, Sarvam, the startup chosen for constructing India’s foundational LLM below the IndiaAI Mission lately unveiled Sarvam-M, a 24-billion parameter open-weights hybrid language mannequin constructed on high of Mistral Small.
The publish Fractal Releases Fathom-R1-14B Reasoning Mannequin on DeepSeek for $499 appeared first on Analytics India Journal.
