Gurugram-based startup KiteFishAI has released NanoI1, a 155-million-parameter open-weight language model trained from scratch in India, on 17 November 2026. The KiteFishAI NanoI1 model is a 311 MB file built for narrow enterprise jobs such as KYC document extraction, complaint classification and support-ticket routing, and can run on-premises or in a private cloud.
The launch adds to a growing list of compact, home-grown models, after NeuroDX’s 240M-parameter Sphota model earlier this month. KiteFishAI pitches NanoI1 as a building block for purpose-built applications rather than a rival to large general-purpose models.

What is KiteFishAI NanoI1?
According to the model card on Hugging Face, NanoI1 is a decoder-only transformer trained on about 1 trillion tokens using eight NVIDIA H100 GPUs. It ships in Base and Instruct versions under the Apache-2.0 licence, so developers and businesses can adapt the weights for their own tasks.
- Size: 155 million parameters, with a 2,048-token context window suited to short documents and forms.
- Architecture: multi-head latent attention, which the company says makes the memory cache about six times smaller than standard attention at the same width.
- Languages: English, Hindi and eight other Indian languages, with multilingual ability the company itself describes as limited.
- Serving: supports vLLM, and is small enough to run on laptops, CPUs, edge boxes and local servers.
Founder and CEO Anuj Gupta said in the company’s announcement: “With NanoI1, we wanted to demonstrate that a compact model, developed from scratch in India, can deliver meaningful performance on specialised tasks while giving organisations greater control over their data and deployment.”

How NanoI1 compares with Gemma-3-270M
In KiteFishAI’s own tests under identical conditions, NanoI1 scored 9.1% on the HumanEval coding benchmark against 2.4% for Google’s larger Gemma-3-270M. It also led on EvalPlus (4.3% vs 1.2%) and GSM8K maths (4.0% vs 1.8%). The model card shows it trails Gemma on commonsense and factual-recall tests, and that Qwen3-0.6B, with about four times the parameters, is stronger on every benchmark listed.
The company is clear that the results do not make NanoI1 better than larger systems in general. Its model card warns against using it for open-ended factual questions, multi-step reasoning, or legal, medical and financial advice without human review.
Built on a small budget with IndiaAI support
Gupta started building models in 2024 while working as a senior AI architect at Accenture and left in August 2026 to run KiteFishAI full-time, according to Analytics India Magazine. He received ₹50 lakh in computing support through an IndiaAI Mission programme, and the company spent about ₹4 lakh training NanoI1. The eight-member team also builds embedding models; it says its Nano-Em1-0.6B-v2.1 ranked ninth on the global MTEB leaderboard in August 2026.
The model joins other Indian efforts to train language models from scratch on modest budgets, such as Murai Labs’ Tamil-first Agaram 2.5B model.
Why it matters for BFSI and enterprise CIOs
For banks, insurers and other regulated firms, a small model that runs inside their own perimeter keeps customer data off external servers and cuts token costs for high-volume, repetitive work. That fits the policy push for indigenous AI models in critical sectors and the rise of on-premise AI servers built for Indian enterprises.
Gupta sees small models working alongside larger ones: a compact model identifies intent, extracts fields or routes a request, while a bigger model handles complex reasoning. He also cautioned that “a smaller language model is not inherently hallucination-free or secure simply because it has fewer parameters”, so access controls, encryption and human review remain necessary.
Key takeaways
- KiteFishAI released NanoI1, a 155M-parameter open-weight model trained from scratch in India, on 17 November 2026.
- It targets KYC extraction, complaint classification and ticket routing, and runs on-premises under an Apache-2.0 licence.
- The startup reports higher coding and maths scores than Gemma-3-270M, but says NanoI1 is not a general-purpose replacement for larger models.
FAQ
What is KiteFishAI NanoI1 used for?
NanoI1 is designed for narrow enterprise tasks such as extracting fields from KYC documents, classifying complaints, routing support tickets and short question answering over supplied documents.
Is NanoI1 free to use?
Yes. The Base and Instruct weights are available on Hugging Face under the Apache-2.0 licence, which allows commercial use and fine-tuning.
Featured image: NVIDIA H100 GPU cards, the type of GPU KiteFishAI used to train NanoI1 (file photo). Photo: Geekerwan via Wikimedia Commons (CC BY 3.0)
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!