2 Sources
[1]
KiteFishAI launches NanoI1, an open-weight AI model for specialised applications
KiteFishAI has introduced NanoI1, targeting developers and enterprises who want to create specialized AI applications. This model is intended to support purpose-built applications instead of competing with large language models. The company highlights the importance of smaller models for businesses
[2]
KiteFishAI Introduces NanoI1, a Compact AI Model to Address Enterprise Data Privacy and Deployment Costs
Trained from scratch on approximately 1 trillion tokens, the 155M-parameter model is designed to bring AI capabilities to enterprise-owned infrastructure As businesses increasingly adopt artificial intelligence, concerns around data privacy, infrastructure costs and dependence on external AI
Share
Copy Link
KiteFishAI, an India-based AI startup, has launched NanoI1, a compact language model with open weights that businesses can run on their own machines rather than through an outside provider. The model, released on 17 November, is pitched at companies that handle sensitive customer data and do not want to send it to an external provider
2
1
.
Source: ET
Anuj Gupta, the company's founder and chief executive, said the challenge for many enterprises lies in deploying AI in ways that meet privacy, infrastructure and cost requirements. "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," he said
2
.NanoI1 has 155 million parameters. It was trained from scratch on roughly 1 trillion tokens using eight H100 GPUs from Nvidia in Gurgaon. The company positions that as the point: this is not a trimmed-down version of someone else's model but one built in India from the ground up
2
.The weights are released under the Apache-2.0 licence, so businesses can adapt the model while keeping their data and infrastructure in-house. The file itself is 311 megabytes, small enough to run on a laptop or a local server, and it can be hosted on-premise or in Indian data centres
2
.It understands English, Hindi and eight other Indian languages, though the company says its multilingual ability is limited. It is built for short, focused jobs: pulling details out of know-your-customer documents, sorting complaints, routing support tickets and processing short documents, with a 2,048-token context window. That window is small, which is another way of saying this is not a general-purpose model and is not pitched as one
2
.In evaluations KiteFishAI ran itself, NanoI1 scored 9.1 percent on HumanEval against 2.4 percent for Google's Gemma-3-270M, a model with nearly twice its parameters; every score is a single-digit percentage. It scored 4.3 percent to 1.2 percent on EvalPlus and by a similar margin on the third test, GSM8K, under the same test settings in each case
2
.The evidence is thin for now: the benchmark scores are self-run and no independent evaluation is reported. The company says the launch aims to put compact open-weight models in the hands of developers, startups and businesses as part of India's growing AI ecosystem
2
.The pitch responds to a problem the company says organisations handling sensitive customer information keep running into: privacy rules, infrastructure costs and dependence on outside model providers. The benchmarks do not settle whether a model this small is good enough for that work. "We believe smaller, purpose-built models can play an important role in making AI more accessible and practical for businesses, particularly in regulated industries," Gupta said
2
.Related Stories
Summarized by
Navi
26 Sept 2025•Technology

04 Nov 2026•Technology

19 Aug 2025•Technology
