Smar NetworkSmar.inSmargrid.comNewsInsuranceRoomsGamesMeasurements

SMAR Daily Industry News

Daily news across AI, technology, finance, insurance, healthcare, automotive, real estate, and government

Facebook Telegram
← Back to AI News

KiteFishAI Builds Tiny AI Models Ranked Among World Best

📅 2026-11-17 | 📰 Analytics India Magazine
#kitefishai #small language models #embedding models #ai governance #enterprise ai

Anuj Gupta, former senior AI architect at Accenture, left the consulting firm in August 2026 to run his Gurugram-based startup KiteFishAI full-time. The company focuses on small, specialised AI models that can perform tasks such as document extraction, complaint classification and ticket routing without the high compute costs of frontier large language models.

KiteFishAI’s flagship compact language model, NanoI1, was trained from scratch on about one trillion tokens using eight NVIDIA H100 GPUs, costing roughly ₹4 lakh in training and development. The 155-million-parameter model occupies 311 MB and is targeted at enterprise use cases like KYC document extraction and short-form question answering. According to the startup, NanoI1 outperformed Google’s Gemma-3-270M on the HumanEval code-generation benchmark, EvalPlus automated test generation, and the GSM8K grade-school math test.

In the embedding space, KiteFishAI released Nano-Em1-0.6B-v2.1, which the company says placed ninth globally on the Massive Text Embedding Benchmark (MTEB) leaderboard in August 2026. Embedding models convert text into numerical vectors, enabling meaning-based retrieval for internal search and retrieval-augmented generation (RAG) applications.

The eight-member team also offers Prism, a model-governance product that evaluates model performance across seven parameters via a single API, though specific parameter details and enterprise deployments were not disclosed. Funding for the venture has been limited to bootstrapping, a ₹50 lakh grant from the IndiaAI Mission, and personal out-of-pocket GPU expenses before securing Azure compute credits from Microsoft.

KiteFishAI’s approach aligns with a broader industry trend toward efficient, smaller models that reduce infrastructure overhead while still handling classification, extraction and routing tasks. Larger models remain available for more complex reasoning, but the startup argues many enterprise workloads do not require them.

The company’s progress demonstrates that with focused engineering and strategic use of government and cloud resources, startups can produce AI components that compete on global benchmarks while keeping costs low for enterprise adopters.

Why it matters

Enterprises can adopt high-performing, low-cost AI models without the expense of massive cloud compute, expanding AI accessibility.

FAQ

When did Anuj Gupta leave Accenture to focus on KiteFishAI?
He left Accenture in August 2026 to work on the startup full-time.

How much did KiteFishAI spend on training NanoI1?
Approximately ₹4 lakh was spent on training and development using eight GPUs.

What ranking did Nano-Em1-0.6B-v2.1 achieve on the MTEB leaderboard?
It ranked ninth globally in August 2026.

What government support did KiteFishAI receive?
The startup received ₹50 lakh in support from the IndiaAI Mission.

Source: Analytics India Magazine. See the original source for full details.