Gurgaon-based AI startup KiteFishAI has launched NanoI1, a 155-million-parameter language model built for specialized enterprise applications. The model focuses on optimizing data privacy, infrastructure efficiency, and deployment costs.
\n\nTrained from scratch on roughly 1 trillion tokens using eight NVIDIA H100 GPUs, NanoI1 features a compact 311 MB file size, enabling seamless deployment on local servers and laptops.
\n\nReleased under the open-source Apache-2.0 license, its open weights allow businesses and developers to customize the model for specific tasks while retaining complete control over their infrastructure and sensitive data.
\n\nNanoI1 is tailored for enterprise workflows like KYC document extraction, complaint classification, customer support ticket routing, and short-document processing. It fully supports hosting within Indian data centers and on-premise environments.
\n\nIn benchmark evaluations conducted by KiteFishAI under identical test conditions, NanoI1 achieved a HumanEval coding score of 9.1% (compared to 2.4% for Google's Gemma-3-270M). It scored 4.3% vs. 1.2% on EvalPlus and 4.0% vs. 1.8% on GSM8K mathematics, despite Gemma-3-270M having nearly double the parameters.
\n\nThe model supports vLLM serving with a 2,048-token context window. Available in both Base and Instruct (chat) versions, it supports English, Hindi, and eight other Indian languages with limited multilingual capabilities.
\n\nThe company emphasized that NanoI1 is meant…
Original source: https://www.cnbctv18.com/technology/