KiteFishAI releases compact AI model for local business workflows
NanoI1 offers adaptable weights under Apache-2.0 and a 311 MB file size. Its company-run coding and math scores do not measure its proposed business uses.
Loading page…
NanoI1 offers adaptable weights under Apache-2.0 and a 311 MB file size. Its company-run coding and math scores do not measure its proposed business uses.
Listen to this story
NanoI1’s strongest evidence so far is also its main limitation: KiteFishAI’s published tests show higher scores than Google’s Gemma-3-270M on three coding and math benchmarks, but provide no accuracy results for the KYC extraction or ticket-routing workflows being pitched. The 155-million-parameter, 311-MB model has an Apache-2.0 license and a 2,048-token context window. Local deployment may suit short, sensitive workflows, but buyers will need task-specific validation; the reported eight-H100 training setup does not establish what hardware inference requires.
KiteFishAI released NanoI1 on November 17, 2026, positioning it as a foundation for specialized applications rather than a general-purpose model replacement.
The company says it trained NanoI1 from scratch on approximately one trillion tokens using eight Nvidia H100 GPUs.
In KiteFishAI’s tests, NanoI1 scored 9.1% on HumanEval, 4.3% on EvalPlus, and 4% on GSM8K, versus Gemma-3-270M’s 2.4%, 1.2%, and 1.8%.
KiteFishAI has launched NanoI1 with a specific pitch: businesses can use a small language model for specialized work while keeping deployment under their control. The India-based startup targets document extraction and customer-support routing, but its published performance comparisons concern coding and mathematics—not those proposed business tasks.
The November 17, 2026 release targets developers, startups and enterprises, including regulated businesses with data-security and deployment requirements. KiteFishAI positions it as a foundation for purpose-built applications rather than a replacement for larger, general-purpose models.
NanoI1 contains 155 million parameters—the learned values that make up the model—and has a reported file size of 311 MB. ETEnterpriseAI says it can run on laptops and local servers, allowing on-premise deployment or hosting in Indian data centres. Greater control over sensitive data and lower infrastructure costs are central to the company's pitch.
The model was trained from scratch on approximately one trillion tokens using eight Nvidia H100 GPUs, according to ETEnterpriseAI. That describes training equipment, not the hardware needed to run a finished application.
Released under Apache-2.0, NanoI1 has weights developers and businesses can adapt. It comes in Base and Instruct versions and supports vLLM, software for serving language models. Its context window—the amount of material it can work with at once—is 2,048 tokens, fitting the launch's emphasis on short documents and forms.
Language coverage includes English, Hindi and eight other Indian languages, though ETEnterpriseAI describes its multilingual capabilities as limited.
Moneycontrol said parameters and detailed benchmark results were not disclosed; ETEnterpriseAI published both. The scores below are KiteFishAI's own evaluations under what it described as identical testing conditions. NanoI1 outscored Google's larger Gemma-3-270M on all three measures.
Those results support a narrow claim: NanoI1 outscored the selected rival on the company's coding and math tests. They are not measurements of complaint classification, identity-document extraction or support-ticket routing. Businesses considering the release are getting an adaptable foundation for those applications, not a demonstrated accuracy result for each proposed workflow.
Loading discussion...
Join the conversation
Explain whether general benchmarks are enough to justify a closer look.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.