Fractal launches Fathom-R1-14B By Kaanchi Chawla - 27 May 2025

Fathom R1 14B

Fractal has launched Fathom-R1-14B, a groundbreaking open-source large language model (LLM) that delivers advanced mathematical reasoning capabilities. Positioned as India’s first step toward building a national-scale reasoning model under the IndiaAI Mission, Fathom-R1-14B is a 14-billion-parameter model that sets new benchmarks in performance and cost-effectiveness.

Developed from Deepseek-R1-Distilled-Qwen-14B, the model has shown superior performance compared to o1-mini and o3-mini, and closely rivals o4-mini (low) on reasoning tasks—all at an affordable post-training cost of just $499.

Fathom-R1-14B is available under the MIT open-source license. Users can explore the model directly on Hugging Face, and access its complete codebase, training recipes, and datasets via GitHub.

“We proposed building India’s first large reasoning model as part of the IndiaAI mission, which includes small, mid-sized, and large (70B parameter) variants,” said Srikanth Velamakanni, Co-founder and Group CEO of Fractal, in a LinkedIn update. “This is just a tiny proof of what’s possible.”

Performance Highlights

The model achieves impressive accuracy on competitive math reasoning benchmarks:

  • AIME-25: 52.71% Pass@1

  • HMMT-25: 35.26% Pass@1

When enhanced with cons@64 (inference-time compute), these scores leap to:

  • AIME-25: 76.7%

  • HMMT-25: 56.7%

The model operates efficiently within a 16K context window, ensuring robust reasoning without extensive compute requirements.

Fractal’s team combined supervised fine-tuning (SFT) with curriculum learning and model merging to optimise performance. A variant, Fathom-R1-14B-RS, further refines these capabilities by integrating reinforcement learning alongside SFT, though at a slightly higher training cost of $967.

“We carefully fine-tuned the model using curated datasets and employed a systematic merging approach to boost accuracy,” the team stated.

This launch follows Fractal’s earlier initiative, Vaidya.ai, a multi-modal AI platform offering free healthcare assistance, showcasing the company’s commitment to socially impactful AI.

In parallel, Sarvam, another startup selected under the IndiaAI Mission, recently introduced Sarvam-M, a 24-billion-parameter hybrid LLM based on Mistral Small, emphasising the growing momentum in India’s open-weight LLM ecosystem.

 

Be a part of Elets Collaborative Initiatives. Join Us for Upcoming Events and explore business opportunities. Like us on Facebook , connect with us on LinkedIn and follow us on Twitter.

"Exciting news! Elets technomedia is now on WhatsApp Channels Subscribe today by clicking the link and stay updated with the latest insights!" Click here!

Related News