KiteFishAI Ships 155-Million-Parameter NanoI1 for Local Enterprise AI
KiteFishAI's NanoI1 is a 311 MB, Apache-2.0 model trained on ~1 trillion tokens in Gurgaon, targeting private local enterprise workflows — with modest benchmarks buyers must test themselves.
By Amara Osei
3 min read

What's News
- KiteFishAI released NanoI1, a 155-million-parameter model, on November 17 under the Apache-2.0 license.
- The model artifact is 311 MB, trained on ~1 trillion tokens using eight NVIDIA H100 GPUs in Gurgaon.
- NanoI1 scored 9.1% on HumanEval versus 2.4% for Google's Gemma-3-270M in vendor-run tests.
- A 2,048-token context window forces applications to split long documents before processing.
- It supports English, Hindi and eight other Indian languages, with capabilities the company calls limited.
KiteFishAI released NanoI1 on November 17, a 155-million-parameter open-weight language model that runs from a 311 MB file on laptops and local servers. The India-based company is pitching it at document extraction, classification and support workflows — not as a replacement for general-purpose AI assistants.
The distinction is the whole pitch. NanoI1's argument isn't that a tiny model can do everything a large system does. It's that a compact, adaptable model may be a more practical foundation for narrowly defined jobs where data handling, hardware requirements and deployment control drive the decision.
What does the model actually offer?
According to KiteFishAI, the company trained NanoI1 from scratch on roughly one trillion tokens using eight NVIDIA H100 GPUs in Gurgaon. The weights ship under the Apache-2.0 license in two versions: Base, a starting point for adaptation, and Instruct, tuned for conversational and task-following applications.
The release lists vLLM serving support and a 2,048-token context window. That limit is a hard boundary. Applications handling long files must select or split their inputs rather than assume an entire document fits in a single request. The company's website points to structured extraction and low-latency tasks among the model's intended uses.
One caveat on the headline number: the 311 MB figure describes the model artifact, not a guarantee that a complete application will need only that much memory.
Which workflows is it built for?
KiteFishAI names know-your-customer document extraction, complaint classification, customer-support ticket routing and short-document processing as potential applications. These are bounded workflows — a system can be scored against a defined output such as a category label or a set of extracted fields.
The model supports English, Hindi and eight other Indian languages, though the company itself characterizes its multilingual capabilities as limited. Buyers should distinguish a language's availability from demonstrated accuracy on their own documents, vocabulary and formatting.
Founder and CEO Anuj Gupta frames the release around enterprise requirements for privacy, infrastructure and cost. In the announcement, he describes smaller, purpose-built models as a way to make AI more practical for organizations that need greater control over where their data and applications reside.
How strong are the benchmark claims?
Vendor numbers are modest. KiteFishAI reports that NanoI1 scored 9.1% on HumanEval versus 2.4% for Google's Gemma-3-270M, in company-conducted evaluations under identical test settings. It also claims 4.3% versus 1.2% on EvalPlus, and 4.0% versus 1.8% on GSM8K.
Those are vendor-reported comparisons, not independent validation. They show an edge over one selected comparison model in the company's own tests — nothing more. The low absolute scores shouldn't be read as reliable general coding or mathematical reasoning, and they don't directly establish accuracy for the document-processing workflows KiteFishAI is targeting.
The more useful enterprise test is whether an adapted version clears a specific workflow's quality threshold. For extraction, that means measuring field-level errors and omissions. For routing, it means measuring classification performance and deciding which uncertain cases stay with a human reviewer.
What does local deployment change — and not change?
NanoI1 supports on-premises deployment and hosting in Indian data centers, per the release. Running locally can eliminate the routine transfer of inputs to an external model provider. It doesn't, by itself, resolve application security, access control, logging or retention requirements.
The practical significance here is the deployment option: an open-weight foundation small enough to evaluate for specialized local applications. Whether that option delivers lower total costs or sufficient accuracy will depend on the workload, integration effort and evaluation results — not parameter count alone.
Source: unite.ai
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,248 articles