With a 240M-parameter model built from first principles and tuned for AMD, Navdyut AI Labs is wagering that relevance and efficiency will beat raw scale.
The prevailing logic in artificial intelligence says bigger wins: more parameters, more GPUs, more capital. Navdyut AI Labs is quietly arguing the opposite. The Guwahati-based lab, co-founded by Dicom Pathak and Lakshya J Bora, has open-sourced a 240-million-parameter foundational model trained entirely from scratch — and its founders say the point isn’t size at all.
“While many labs focus on bigger models, we focus on relevant models. The biggest challenge in the industry right now boils down to control, inference cost, and the global chip shortage. By training from scratch, we dictate exactly what goes into the AI’s diet.” — Dicom Pathak, Co-founder, Navdyut AI Labs
For Navdyut, “from scratch” is not a marketing line. Most startups take an open-weight model from Meta, Google or another large lab and fine-tune it. Navdyut built its own tokenizer, designed its own training pipeline and worked through the mathematics of Maximal Update Parameterization (muP) and gradient control, while staying close to Chinchilla-optimal scaling laws. The result is the latest in a Navdyut model family that has grown steadily from an initial 15 million parameters.
The hardware wager
The most strategic choice Dicom Pathak and Navdyut have made may be the least visible one: the hardware. The AI industry runs overwhelmingly on Nvidia’s CUDA platform, a convenience that has become a liability as chip supply tightens. Navdyut is engineering its models for AMD inference instead.
“We are engineering these models specifically for AMD inference to escape the CUDA bottleneck, enabling seamless transitions and significantly cheaper real-world deployment.” — Dicom Pathak, Navdyut AI Labs
If the bet works, Navdyut’s models could be deployed more cheaply and with less exposure to the supply constraints that shape who can — and cannot — build AI today.
Inside the Navdyut roadmap
Navdyut treats the 240M model as proof that its infrastructure works, not as an end product. Next on the Navdyut roadmap are 480M and 960M models. Pathak’s team sees 960M as the threshold for real production use, bringing agentic tool-calling and Domain Adaptive Pre-Training, which tailors a model deeply to one domain.
The bigger idea is “Modular Agentic Foundational Models” — compact, specialised models that slot together like Lego pieces instead of one giant system that tries to do everything. Navdyut’s target is for a focused 960M model to rival a 1.5B general-purpose model while cutting inference compute by up to half, making on-device deployment practical. After that, Navdyut plans to move into the 1.5B–8B range, which it sees as the balance point between low compute cost and enterprise capability.
The takeaway
Navdyut is bootstrapped, small and outside India’s venture-funded metro ecosystem, yet it places itself among a single-digit group of Indian organisations training true foundational models from scratch. Navdyut’s AMD-first, modular strategy still has to prove itself at larger scales. For now, Dicom Pathak and Navdyut have put a real, downloadable model on the table — and a clear argument that relevance can compete with scale.
Navdyut’s 240M model and the rest of the Navdyut model family are available on Hugging Face at huggingface.co/dicompathak. More about Navdyut AI Labs is at navdyut.com.