Fractal launches Fathom-R1-14B, a reasoning-focused open-source AI model

Fractal, a Mumbai-based AI company, has launched Fathom-R1-14B, an open-source large language model with 14 billion parameters.

This release is part of Fractal’s initiative under the Government of India’s IndiaAI Mission, aimed at building India’s first large reasoning models.

What is Fathom-R1-14B?

Fathom-R1-14B is designed to excel in complex reasoning and mathematical problem-solving tasks. The model is based on the DeepSeek-R1-Distilled-Qwen-14B architecture, which has been improved via supervised fine-tuning, curriculum learning, and model integration approaches. The post-training expense of Fathom-R1-14B was only $499, demonstrating a cost-effective strategy for creating high-performance AI.

The model has demonstrated strong results on challenging benchmarks:

• 52.71% accuracy on AIME-25 (Pass@1)

• 35.26% accuracy on HMMT-25 (Pass@1)

These scores increased to 76.7% and 56.7%, respectively, with inference-time improvements (cons@64).

Additionally, utilizing reinforcement learning, Fractal unveiled a version called Fathom-R1-14B-RS. With a training expense of $967, this version produced similar outcomes.

Open Access and Future Plans

The MIT license, with the source code available on GitHub and model access given on Hugging Face, allows public access to Fathom-R1-14B.

The launch aligns with the Government of India's ₹10,372 crore IndiaAI compute initiative, which seeks to make the country a center for AI R&D worldwide. Fractal is one of several contributors, including Sarvam AI, which just released Sarvam-M, a hybrid language model with 24 billion parameters.

"As part of the IndiaAI mission, we suggested constructing India's first large reasoning model. We suggested creating three models—a little one, a medium one, and a big one with 70 billion parameters” -stated Fractal CEO Srikanth Velamakanni in a LikedIn post.

The introduction of Fathom-R1-14B marks a significant step in India's AI ecosystem, demonstrating the feasibility of developing cost-effective, open-source AI models at scale. This initiative is expected to drive innovation across academic research, enterprise applications, and public sector deployments, reinforcing India's commitment to digital sovereignty and democratized AI development.