Originals · Startup · Startups & Fintech
KiteFishAI releases NanoI1, a 155 million parameter open model small enough to run on a laptop
The Gurgaon startup trained the model from scratch for narrow enterprise jobs such as KYC extraction and ticket routing, and published the weights under Apache 2.0.

Gurgaon based KiteFishAI has launched NanoI1, a compact open-weight language model with 155 million parameters, the company announced on 17 November 2026. The Economic Times, Moneycontrol and CNBC TV18 reported the launch, and both the Base and Instruct versions are live on the company's Hugging Face page under the Apache 2.0 licence.
The pitch is the opposite of the frontier race. KiteFishAI says NanoI1 is meant to be a foundation for purpose-built applications, not a rival to large general models. CNBC TV18 reported the model was trained from scratch on roughly 1 trillion tokens using eight NVIDIA H100 GPUs, and the weights file is about 311 MB, small enough for local servers and laptops.
The target users are businesses in regulated sectors that want to keep data inside their own walls. The company lists jobs like KYC document extraction, complaint classification, customer support ticket routing and short document processing, and says the model can be hosted in Indian data centres or fully on premise. It supports vLLM serving with a 2,048 token context window.
On language, the model card lists English, Hindi, Bengali, Marathi, Tamil, Telugu, Gujarati, Kannada, Malayalam and Punjabi, though the company describes the multilingual ability as limited. In its own benchmarks run under identical conditions, KiteFishAI says NanoI1 scored 9.1% on HumanEval against 2.4% for Google's Gemma 3 270M, which has nearly twice the parameters. These are company figures and have not been independently tested.
NanoI1 is the first model family the company built end to end on its own architecture, tokenizer and training stack, according to its website. It sits alongside other open models from the startup, including a 10 billion parameter reasoning model, a 2 billion parameter maths model and a 0.6 billion parameter embedding model aimed at private search and retrieval.