Article complete

He Left Accenture to Build an Embedded SLM—Now It Ranks Among the World’s Best

Anuj Gupta founded KiteFishAI to develop compact language and embedding models for enterprise applications, AI governance, and multimodal technologies.

Small language model

Key takeaways

  • Anuj Gupta founded KiteFishAI to create compact language and embedding models tailored for enterprise applications.
  • He challenges the notion that larger models are always superior, advocating for specialised systems for specific tasks.
  • The startup focuses on various AI technologies, including retrieval-augmented generation and model governance with its product Prism.
  • Gupta transitioned from Accenture to full-time entrepreneurship, overcoming financial challenges and securing government support for computing resources.
  • KiteFishAI has gained recognition for its innovative approach in the competitive AI landscape.

Takeaways drafted with AI from the article text.

As enterprises adopt AI and move from pilots to production, AI finops is increasingly seeping into deployment decisions. While frontier large language models can write production code and run multi-step agentic workflows, they’re expensive, especially when token costs stack up.

Anuj Gupta, Founder and CEO of KiteFishAI, believes many enterprise applications do not require large models. Tasks such as extracting information from documents, classifying customer complaints, and routing support tickets can be handled by smaller, specialised systems.

“We wanted to challenge a common assumption in the AI industry that better intelligence always requires larger models,” Gupta tells AIM.

The Gurugram-based startup is developing small and large language models, embedding models and enterprise AI technologies. Its work also spans retrieval-augmented generation (RAG), AI agents, scientific AI, and model governance through a product called Prism.

From AI Architect to Founder

Gupta began experimenting with model development in 2024 while working as a senior AI architect at Accenture, where he spent nearly five years. Access to computing infrastructure, however, became an early obstacle. GPU resources were expensive for an individual developer attempting to build and train models independently.

Gupta approached Microsoft for access to Azure computing resources at a lower cost. He used the infrastructure to develop his first model, which he released in January 2026. He later approached the IndiaAI Mission for GPU access and received additional computing support through a government programme. Through the programme, he received ₹50 lakh in government support.

He left Accenture in August 2026 to work on the startup full-time. However, the transition also involved personal financial pressure. “I was spending money from my pocket to continue using the GPUs,” he recalls.

The startup is currently bootstrapped.

That experience has influenced the company's approach to developing AI systems with limited resources. KiteFishAI operates with a small engineering team. For NanoI1, one of its compact language models, the company spent approximately ₹4 lakh on training and development using eight GPUs.

Building an Enterprise AI Portfolio

KiteFishAI's work extends beyond individual language models. Its areas of development include embedding models, large language models, RAG systems, AI agents, and scientific AI.

Embedding models represent text as numerical vectors, allowing applications to retrieve information based on meaning rather than relying only on exact keyword matches. They are used in enterprise search and RAG applications, where AI systems need to retrieve relevant information from internal documents and databases.

One of KiteFishAI's models in this category is Nano-Em1-0.6B-v2.1, which the company says ranked ninth globally on the Massive Text Embedding Benchmark (MTEB) leaderboard in August 2026. MTEB evaluates embedding models across different text retrieval and representation tasks.

The eight-member company is also developing Prism, a model governance product that evaluates model performance across seven parameters through a single API. The product looks to help organisations govern model use and identify potential misuse. However, the startup didn’t share specific details on the seven parameters and enterprise deployments.

Liquid AI is another player in the smaller-model space. The company develops Liquid Foundation Models (LFMs), a family of AI models designed to be efficient across different computing environments, including edge devices, personal computers, and enterprise infrastructure. 

The broader strategy is to develop components that can work together in enterprise AI applications. Smaller models can handle classification, extraction and routing, while embedding systems support information retrieval. Larger models can then handle tasks that require more complex reasoning.

Why Smaller Models

KiteFishAI believes enterprises do not need large language models for every operation. Although larger systems can handle complex reasoning and a wider range of tasks, they can also require greater computing resources and introduce additional infrastructure and governance considerations.

NanoI1 illustrates the company's approach. The 155-million-parameter language model was trained from scratch on approximately 1 trillion tokens using eight NVIDIA H100 GPUs. It has a file size of 311 MB and is designed for specialised applications such as KYC document extraction, complaint classification, support-ticket routing, and short-form question answering.

The startup claims that NanoI1 scored higher than Google's Gemma-3-270M on HumanEval code generation benchmark, EvalPlus on automated test generation, and grade school math-focused GSM8K. The company has clarified that these results do not make the model universally better than larger systems. Its applications are narrower, focusing on tasks where compact size and deployment flexibility are important.

The model is available under the Apache-2.0 licence in Base and Instruct versions. It supports English, Hindi, and eight additional Indian languages, although the company acknowledges that its multilingual capabilities remain limited.

For enterprises, NanoI1 can be deployed on-premises or in a private cloud, allowing inference to take place within infrastructure controlled by the organisation.

“A smaller language model is not inherently hallucination-free or secure simply because it has fewer parameters,” he explains. Enterprises still need safeguards such as access controls, encryption, application-level permissions, and human review for sensitive decisions.

Moving Towards Multimodal AI

Beyond text-based models, KiteFishAI plans to expand into voice, audio, and video. Gupta identifies voice-based models as an area of current focus, with plans to embark on multimodal development over the next two years.

India remains its primary focus, and improving language capabilities is part of its development direction. The company also plans to explore longer context windows, specialised models for industry-specific requirements, and integrations with enterprise applications.

Gupta expects smaller models to work alongside larger reasoning systems rather than replace them. In an enterprise workflow, a compact model could identify intent, extract information or route a request, while a larger model handles tasks requiring broader reasoning.

“See if your task—if it's a particular task that you want—go and use that SLM. Get your task done. Pay less for that particular task,” he advises.

The company is also exploring RAG and agentic workflows, in which different models can be assigned specific responsibilities instead of relying on a single system for every operation.

For KiteFishAI, the next step will be translating its model development work into wider enterprise adoption.

Published

Corrections: AIM corrects errors of fact and notes the change on the story. Spotted one? Tell the editors. Read our editorial standards.

More in this story

enterprise AI · 5 stories
  1. HCLTech Bags $1.14 Bn Mega AI Deal from Fortune Global 50 European Firm
  2. Can Enterprise AI Agents Learn From Their Own Mistakes? A Bengaluru Startup Thinks So
  3. As AI Agents Leave Sandboxes, Enterprises are Hitting the Governance Panic Button
  4. The AI Slowdown Narrative Doesn’t Reflect Enterprise Reality
  5. You are here: He Left Accenture to Build an Embedded SLM—Now It Ranks Among the World’s Best
About the author

Contributor

Kileuna is a journalist covering the real-world impact of artificial intelligence at the intersection of tech, policy, and society.

All stories
Belamy · Daily

The day’s AI news, before your first meeting.