wasl.tech

KiteFishAI NanoI1: A 155M-Parameter AI Model Built for Banks, Not Chatbots

The Morning a Bank Clerk Met a 311 MB Brain

Picture a compliance officer at a mid-sized Indian bank, staring at a tower of scanned identity documents on a Monday morning. Every page has to be checked, every name matched, every address pulled out by hand before the customer can open an account.

It is slow, repetitive work, and the bank cannot simply send those documents to a giant cloud chatbot. The papers hold passports, signatures and account numbers. The law says they stay inside the building.

That is exactly the problem a small startup from Gurugram set out to solve. On November 17, 2026, KiteFishAI released NanoI1, an artificial intelligence model so small it fits on a laptop — and so focused it only wants to do a handful of jobs, but do them well.

NanoI1 is not trying to be the next ChatGPT. It cannot write your novel or plan your holiday. What it can do is read a form, pull out the right fields, sort a complaint, and route a support ticket — all without a single byte of customer data ever leaving the company’s own servers.

What Is KiteFishAI NanoI1?

KiteFishAI NanoI1 is a compact, open-weight language model built for specialised enterprise work. It has just 155 million parameters, making it a fraction of the size of the flagship models that dominate the headlines.

The model comes in two versions, Base and Instruct, and its weights are released under the Apache-2.0 licence. That means developers and businesses can download it, fine-tune it on their own data, and use it commercially, according to The Economic Times.

At 311 MB, the whole model is smaller than a short phone video. It runs on ordinary laptops and local servers, and it supports vLLM serving for teams that want to run it at scale.

The company describes NanoI1 as a foundation for purpose-built applications rather than a general-purpose chatbot. It is the difference between a Swiss Army knife and a scalpel: one does a little of everything, the other does one job precisely.

Trained From Scratch in India

NanoI1 was trained from scratch on approximately one trillion tokens, using eight NVIDIA H100 graphics processors in Gurgaon. That is a modest setup by the standards of today’s AI labs, where training runs can involve thousands of chips.

The training happened entirely in India, and the company says that matters. For banks, insurers and other regulated businesses, where data lives and who controls it is not a technical detail — it is a compliance requirement.

KiteFishAI also built the model to speak India’s languages. Alongside English, NanoI1 supports Hindi and eight other Indian languages, with limited multilingual ability. That is a practical choice for a country where customer documents and support tickets arrive in many tongues.

The model works with a context window of 2,048 tokens, which is short by modern standards. It suits short documents, forms and structured workflows — exactly the material it is meant to process.

Small by Design: Why 155 Million Parameters Is the Point

Most AI news in 2026 is about models getting bigger. KiteFishAI is betting the other way. A 155-million-parameter model costs far less to run, needs far less hardware, and can live entirely on a company’s own machines.

That changes the economics for smaller firms. A business that could never afford a rack of expensive GPUs can run NanoI1 on equipment it already owns. And because the model never has to phone home to a cloud provider, sensitive data never crosses the internet.

This is part of a wider shift in the industry. Open-weight models have been moving in both directions this year — some, like Meituan’s LongCat-2.5, pack trillions of parameters, while others chase efficiency at the small end. NanoI1 plants its flag firmly at the small end.

The trade-off is real, and KiteFishAI does not hide it. A tiny model cannot know as much as a giant one. The question is whether it needs to — and for reading a KYC form, the answer may well be no.

How NanoI1 Stacks Up Against Gemma-3-270M

In its own tests, run under identical conditions, KiteFishAI says NanoI1 beat Google’s larger Gemma-3-270M on coding and maths benchmarks. On HumanEval, a coding test, NanoI1 scored 9.1% against Gemma’s 2.4%. On EvalPlus, another coding benchmark, it scored 4.3% against 1.2%.

On GSM8K, a mathematics test, NanoI1 reached 4.0% while Gemma-3-270M managed 1.8%. These are low scores in absolute terms — neither model is a maths star — but the gap is notable given that Gemma has nearly twice the parameters.

The company is honest about where its model falls short. Its model card shows NanoI1 trailing Gemma on commonsense and factual-recall tests. And Qwen3-0.6B, with about four times the parameters, is stronger on every benchmark listed.

That candour is refreshing in an industry where launch announcements usually read like victory laps. KiteFishAI is not claiming to have built a better brain. It is claiming to have built a cheaper, more practical one for a specific kind of work.

Built for Banks and Regulated Industries

NanoI1’s target jobs are unglamorous but valuable. The company lists KYC document extraction, complaint classification, customer support ticket routing and short-document processing as its intended uses.

These are the chores that eat up staff hours in banking, insurance and financial services — collectively known as BFSI. A model that can read a scanned identity card and pull out the name, date of birth and address, reliably and privately, is worth more to a bank than a model that can write poetry.

Regulated industries have been the slowest to adopt AI precisely because of data rules. Sending customer records to an outside cloud service raises legal and security questions that many compliance teams would rather not answer.

A model that runs on-premises or in a private cloud sidesteps that problem. The data stays where it started, the model runs where the company can see it, and the audit trail is simple. For a bank’s chief compliance officer, that may matter more than any benchmark score.

What NanoI1 Cannot Do

Small models come with clear limits, and NanoI1’s own model card spells them out. It warns against using the model for open-ended factual questions, multi-step reasoning, or legal, medical and financial advice without human review.

That is an important warning. A 155-million-parameter model does not have the knowledge baked in to answer general questions reliably, and in regulated industries a wrong answer is not just embarrassing — it can be expensive.

The model is a tool for narrow, supervised tasks, not a replacement for human judgement. The companies that get the most from it will be the ones that put people in the loop: the model drafts, extracts and sorts, and a human checks and signs off.

This is also why the comparison with giant models misses the point. NanoI1 is not competing with the likes of Qwen4-Max or the flagship systems from the big labs. It is competing with the manual work that nobody wants to do, and with the expensive cloud bills that come with sending every document to a giant model.

Why Small Models Are Having a Moment

NanoI1 arrives at a time when the AI industry is quietly rediscovering small. After years of racing to build ever-larger models, companies are asking what they actually need — and finding that many jobs do not require a trillion parameters.

Open-weight releases have made this possible. When anyone can download a model’s weights and adapt them, the barrier to building a specialised tool collapses. We have seen the same pattern with image models like MiniMax H3-Image, which put serious capability into a downloadable package.

For India, NanoI1 is also a small piece of a bigger story. The country is building its own AI ecosystem, from startups to government-backed compute infrastructure, and locally trained models are part of that push.

KiteFishAI says it wants to contribute to that ecosystem by making compact models available to developers and businesses. Whether NanoI1 finds a real audience will depend on how well it performs on real documents, in real offices, under real compliance rules — not on benchmark tables.

But the direction is clear. AI is no longer only about the biggest model in the world. Sometimes, the most useful model is the one small enough to fit on the machine already sitting on your desk.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top