Models6↗open-weight on Hugging Face Downloads94k+↗downloads on Hugging Face

kitefish / models

$open-source model registry

01Reasoning-L1-10Bopen weights↗02Nano-Math-2B-v2open weights↗03Nano-Em1-0.6B-v2.1open weights↗04Nano-Math-2B-v1open weights↗

no gate / no waitlistdownload →

Hindi

Small Language Models

built for your industry

We build compact, high-performance language models for BFSI, healthcare, pharma, and beyond. Open-source base models, on-prem deployment, built for India and the world.

Request Early Access→View Models

On-prem / air-gapped / cloud private

Open-source models

Built small.Reasoning big.

Compact models across OCR, embeddings, reasoning, and speech workloads, ready for domain fine-tuning and private deployment.

Math & Science SLM

Nano-Math-2B-v2

Model

Scaled math reasoning model fine-tuned for high accuracy on production scientific and mathematical workloads.

2BSize

1.5k+Downloads

CurrentUpdated

Quick start / 2B

Nano-Math-2B-v2

Math & Science SLM

Open HF

from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="KiteFishAI/Nano-Math-2B-v2",
)

prompt = "Solve step by step: Calculate the derivative of f(x) = x^3 * e^(2x)."
response = pipe(prompt, max_new_tokens=200)

print(response[0]["generated_text"])

Quick start / 2B

Nano-Math-2B-v2

Math & Science SLM

Open HF

from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="KiteFishAI/Nano-Math-2B-v2",
)

prompt = "Solve step by step: Calculate the derivative of f(x) = x^3 * e^(2x)."
response = pipe(prompt, max_new_tokens=200)

print(response[0]["generated_text"])

Quick start / 10B

Reasoning-L1-10B

MoE reasoning

Open HF

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "KiteFishAI/Reasoning-L1-10B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

prompt = "Reason step by step: if a train covers 180 km in 3 hours, what is its speed?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=180)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Industry fine-tuning

One model family.Every domain.

Base models are the foundation. We adapt them into industry language, workflows, and edge cases with pilot partners.

BFSI workflows

Banking & Finance

Compliance document analysis, credit risk reasoning, regulatory Q&A, and financial report summarisation tuned for Indian BFSI workflows.

Risk reviewCredit memoRegulatory Q&A

Talk to us →View models

Why KiteFish

Small, domain-specific,domain-ready.

Four design choices shape every KiteFish model: private deployment, compact size, industry depth, and open foundations.

01 / Small by design

Run entirely on-premise

Your data never leaves your infrastructure. KiteFish models are built to deploy inside your environment — air-gapped, compliant, and fully under your control.

Learn more →

02 / Right-sized

Small models, real intelligence

From 0.6B to 10B parameters, our models are designed to run efficiently on hardware you already own, without sacrificing domain accuracy for size.

Learn more →

03 / Domain depth

built for your industry

General-purpose LLMs are trained for everything. KiteFish models are built to be fine-tuned deep into your specific domain — the language, the tasks, the edge cases.

Learn more →

04 / Open by default

Open-source foundation

All our base models are released on Hugging Face. Download, evaluate, and fine-tune freely. Enterprise support and custom deployments available on request.

Learn more →

The Team

Practitioners building for practitioners.

Founded by senior practitioners from Accenture's financial services and technology practice — with hands-on experience delivering AI at enterprise scale across India's largest organisations.

Combined patents across financial technology and pharmaceutical domains. Backed by Microsoft for Startups with active Nvidia H100 infrastructure and a published research paper on model compression.

Microsoft for StartupsAccenture AlumniMulti-domain patentsNvidia H100LM_eval Benchmarked

Nano-Math-2B-v2 Released

Math & scientific reasoning model

Nano-Em1-0.6B-v2.1 Released

MTEB & retrieval embedding model

Reasoning-L1-10B Released

Flagship MoE reasoning model · 10B

Research Paper Published

MoE compression methodology

Microsoft for Startups Grant

Nvidia H100 Infrastructure active

Nano-Math-2B-v1 Released

Compact edge math model

Pitching Industry Partners

BFSI · Healthcare · Pharma · Legal

Founders story

Why we are building small language models.

KiteFishAI started with a practical belief: advanced AI should be usable by teams that cannot send sensitive data outside their own systems. Our founders are building smaller, domain-ready models that enterprises can evaluate openly and deploy with control.

Built for teams that need AI inside their own infrastructure.

Focused on compact models that can be evaluated, owned, and deployed privately.

Designed around regulated workflows where accuracy, control, and cost all matter.

Meet KiteFishAI

// Model access

Build with KiteFishAI models.

OCR, embeddings, reasoning, speech, vision-language, and math models are ready for evaluation and domain training.

Industry-tuned

BFSI, healthcare & pharma

Base models

Available on Hugging Face

Pilot access

Co-develop domain models

Request Early Access→Download Base Models