In a significant move underscoring India's growing ambitions in artificial intelligence, Mumbai-headquartered Fractal has announced the launch of its new open-source reasoning model, Fathom-R1-14B, featuring 14 billion parameters.
This launch is part of the company's efforts to establish India's first large reasoning model, under the umbrella of the "IndiaAI Mission."
The company clarified via a LinkedIn post that this model was specifically designed for tasks requiring high capabilities in mathematical and general reasoning.
Fathom-R1-14B was built on the Deepseek-R1-Distilled-Qwen-14B framework and boasts a context length of up to 16,000 tokens, a feature that allows it to process longer inputs within a single prompt.
The model demonstrated superiority in self-consistency tests, which measure its ability to provide coherent answers across multiple attempts, outperforming other renowned models like o3-mini and o1-mini.
You can access the model from its Hugging Face page, and you can also try it for free on this space.
The Race for AI Reasoning Models
The launch of Fathom-R1-14B coincides with a significant shift in global AI development priorities.
The sector is transitioning from models focused on fluent text generation towards those capable of logical and structured reasoning.
This transition marks the new frontier in AI evolution, with leading companies like OpenAI, with its o1/o3 model series, and Nvidia, with its Llama Nemotron family, developing specialized reasoning models that excel in mathematics, programming, and step-by-step problem-solving.
To provide a clearer picture of the scale of investment in these capabilities, the DeepSeek-R1 model, which formed the basis for Fractal's new model, boasts a total of 671 billion parameters, with 37 billion parameters activated per token.
Undoubtedly, the performance improvements resulting from this trend are substantial, as these reasoning-focused models can show accuracy improvements of up to 20% in complex tasks compared to base models, which explains Fractal's emphasis on Fathom's performance in advanced mathematics benchmarks.
Industry analysts anticipate this competitive landscape will intensify throughout 2025, although some experts caution that innovation in reasoning models might slow down as it approaches fundamental scaling limits.
India Reinforces Its Steps Towards AI Self-Sufficiency
Fractal's development of the Fathom-R1-14B model – hailed as "India's first large reasoning model" – represents a significant milestone within the country's broader AI ambitions.
It also aligns with the government-backed "IndiaAI Mission."
In this context, India has earmarked substantial resources for developing local AI capabilities, including investments in GPU infrastructure amounting to 18,693 units, which are crucial for training advanced models like Fathom.
The timing here is strategic, as India plans to launch its first comprehensive indigenous foundational AI model in 2025, positioning the country to compete with global AI leaders like the US and China.
This drive towards AI self-sufficiency is fueled by economic considerations, with projections indicating that AI could contribute up to $1 trillion to India's economy by 2035.
A significant portion of this figure is expected to come from generative and reasoning AI applications.
Fractal's approach reflects a notable 'collaborative self-reliance' strategy across India's AI ecosystem, which builds upon existing global technologies while developing distinctive capabilities tailored to local needs.
Open-Source AI Expands Access to Reasoning Capabilities
Fractal's decision to release Fathom-R1-14B as an open-source model underscores the growing trend towards democratizing access to advanced AI technologies, especially in the realm of reasoning models.
The model's low post-training cost – a mere $499 USD – illustrates how open-source development lowers financial barriers to creating sophisticated AI models, enabling smaller organizations and researchers to participate in the latest advancements in this field.
This approach mirrors other recent releases like Qwen QwQ, which achieved a 90.6% score on the MATH-500 benchmark while maintaining transparency and accessibility.
Thus, open-source reasoning models unlock innovation across various sectors by allowing customization for specific use cases without the prohibitive costs associated with proprietary systems, although they still face challenges in matching the performance of the largest proprietary models in some scenarios.
Nvidia's recent initiative to make its AI tools and datasets openly available further reinforces this trend towards expanding the beneficiary base, suggesting that competitive reasoning capabilities will become increasingly accessible to developers worldwide.
In a LinkedIn comment, Srikant Velamakanni, CEO of Fractal, stated: "We proposed building India's first large reasoning model as part of the IndiaAI Mission. We proposed building three models (a small one, a medium-sized one, and a large one with 70 billion parameters)."
He added: "This is just a small demonstration of what can be achieved."
It is worth noting that Fractal also unveiled a separate version (Fathom-R1-14B-RS) that achieved similar results using a combination of reinforcement learning and supervised fine-tuning, at a cost of $967.
Last year, the company launched Vaidya.ai, a multimodal AI platform designed to provide free and accessible healthcare assistance.
