KiteFishAI Launches NanoI1, a 155M Open-Weight AI Model
The Gurugram startup's compact model is built for tasks like KYC document extraction and runs on local servers.
Key Takeaways
- KiteFishAI has launched NanoI1, an open-weight model with 155 million parameters.
- It was trained from scratch on about 1 trillion tokens using 8 NVIDIA H100 GPUs.
- The weights are under the Apache-2.0 licence and the model file is about 311 MB.
- KiteFishAI said it suits KYC document extraction, complaint classification and ticket routing.
- In KiteFishAI's own tests, it scored 9.1% on HumanEval against 2.4% for Gemma-3-270M.
Table of Contents 7 sections
KiteFishAI, an AI startup based in Gurugram, has launched NanoI1, a compact open-weight language model with 155 million parameters. The company said the model is meant to be a foundation for specialised applications, not a general-purpose rival to large language models.
NanoI1 was trained from scratch on about 1 trillion tokens using 8 NVIDIA H100 GPUs. Its weights are released under the Apache-2.0 licence, and the model file is about 311 MB, so it can run on laptops and local servers. It comes in Base and Instruct (chat) versions.
KiteFishAI NanoI1 at a glance
| Detail | Value |
|---|---|
| Company | KiteFishAI |
| Based in | Gurugram |
| Founder and CEO | Anuj Gupta |
| Model | NanoI1 (Base and Instruct versions) |
| Size | 155 million parameters, model file of about 311 MB |
| Training | About 1 trillion tokens, 8 NVIDIA H100 GPUs |
| Licence | Apache-2.0, open weights |
| Context window | 2,048 tokens |
| Languages | English, Hindi and eight other Indian languages (limited) |
| Announced | November 17, 2026 |
KiteFishAI built NanoI1 for KYC, complaints and support tickets
KiteFishAI said NanoI1 is designed for focused enterprise work such as KYC document extraction, complaint classification, customer support ticket routing and short-document processing. It can be run on-premise or hosted in Indian data centres, so companies can use it without routinely sending sensitive data to outside model providers. The model supports vLLM serving, and its 2,048-token context window suits short documents, forms and structured workflows.
The model card on Hugging Face says NanoI1 is built for low-cost, on-premise use in Indian regulated industries such as banking, insurance and financial services. The same card notes that its ability in Indian languages is limited, since they make up a small share of its training data.
How NanoI1 scored against Gemma-3-270M in KiteFishAI's tests
In evaluations run by KiteFishAI under the same test settings, NanoI1 was compared with Google's Gemma-3-270M. These are the company's own results.
| Benchmark | NanoI1 | Gemma-3-270M |
|---|---|---|
| HumanEval (coding) | 9.1% | 2.4% |
| EvalPlus (coding) | 4.3% | 1.2% |
| GSM8K (maths) | 4.0% | 1.8% |
Anuj Gupta on compact models for regulated industries
"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. We believe smaller, purpose-built models can play an important role in making AI more accessible and practical for businesses, particularly in regulated industries."
What KiteFishAI says it wants NanoI1 to do
KiteFishAI said that with NanoI1 it aims to contribute to India's AI development ecosystem by making compact, open-weight models available to developers, startups and enterprises. The company said the model is intended for purpose-built applications, particularly in regulated industries.
Where KiteFishAI stands after the NanoI1 launch
KiteFishAI has released NanoI1, a 155 million parameter model trained on about 1 trillion tokens, as open weights under the Apache-2.0 licence. It is available in Base and Instruct versions for tasks such as KYC document extraction and support ticket routing. The company has positioned it as a foundation for specialised applications that can run on a company's own servers.
More news
Frequently Asked Questions
What is KiteFishAI?
What is NanoI1?
How big is KiteFishAI's NanoI1 model?
How was NanoI1 trained?
Under which licence did KiteFishAI release NanoI1?
What tasks is KiteFishAI's NanoI1 meant for?
How can companies deploy KiteFishAI's NanoI1?
How long a text can KiteFishAI's NanoI1 handle?
Which languages does KiteFishAI's NanoI1 support?
How did NanoI1 perform in KiteFishAI's benchmark tests?
Who founded KiteFishAI?
What versions of NanoI1 has KiteFishAI released?
More in AI
Google Rolls Out Tasks in Gemini for Free in India
Google has started rolling out Tasks in Gemini, its personal AI assistant, free of charge in India.
Sarvam AI Opens Sarvam Token Factory to All Developers
Sarvam AI has opened Sarvam Token Factory, its India-hosted AI inference service, to everyone.
SquadStack Launches Maya, a Sales LLM Built and Hosted in India
SquadStack has launched Maya, a large language model it built only for live sales calls.
