KiteFishAI bets on compact models for enterprise tasks
Gurugram-based KiteFishAI is building smaller, specialised models for enterprise workloads that do not require large reasoning systems. Founder and CEO Anuj Gupta left Accenture in August 2026 to pursue the bootstrapped startup full-time after developing models independently. The company received ₹50 lakh in government support through a programme providing access to computing resources.
Its NanoI1 language model has 155-million-parameter architecture and was trained from scratch on approximately 1 trillion tokens using eight NVIDIA H100 GPUs. Training and development cost approximately ₹4 lakh. Designed for document extraction, complaint classification, support-ticket routing and short-form question answering, it can run on-premises or in a private cloud. KiteFishAI claims benchmark advantages over Google’s Gemma-3-270M on selected coding and maths evaluations, while acknowledging that these results do not establish broader superiority. Separately, the company says its Nano-Em1-0.6B-v2.1 embedding model ranked ninth globally on the Massive Text Embedding Benchmark leaderboard in August 2026.
The startup is also developing Prism, a governance product that evaluates models across seven parameters through a single API, but has not disclosed those parameters or enterprise deployments. Its roadmap includes voice and multimodal systems, improved Indian-language capabilities and industry-specific models. Gupta cautions that compact models still require security controls and human review, positioning them as complements to larger systems rather than replacements.