Open AI Models for Nepal: Nepali Language and Low Connectivity

In short: Open-weight AI models can be downloaded and run on your own phone, laptop or server, which matters in a country where about 56% of people are online and mobile coverage is patchy in rural areas. The open question is language: published tests show open models scoring lower in Nepali than in English, and we found no Nepali benchmark for Gemma 4 yet.

Key Takeaways

• Google says the smaller Gemma 4 models are built to run locally on laptops and mobile devices, under the Apache 2.0 license.
• Nepal has far more mobile connections than people online: DataReportal counts 32.4 million connections but 16.6 million internet users, and 77% of people live in rural areas.
• On IndicParam, a 2026 benchmark, Gemma 3 scored roughly 29% to 39% in Nepali, and the best model tested scored 52%.
• A small community adapter shows Gemma 4 can be steered to answer in Devanagari, but that measures script, not quality.
• Before building on any open model, test it on your own Nepali documents.

The Evidence

Three kinds of evidence matter here: what open models can do, how well they handle Nepali, and what connectivity looks like on the ground.

What the models can do. Google's April 2, 2026 announcement describes Gemma 4 as open models under the Apache 2.0 license, with edge-sized versions (E2B and E4B) and support for more than 140 languages. The Gemma 4 model card says the models were pre-trained on 140+ languages and support 35+ languages out of the box, and that the smaller ones are designed for efficient local execution on laptops and mobile devices. It also lists a 12B variant not in the launch post. Its one multilingual benchmark, MMMLU, gives 76.6% for E4B and 67.4% for E2B, but it is not a Nepali test.

How well they handle Nepali. The closest independent evidence is IndicParam, a benchmark from researchers at the Indian Institute of Management Indore (January 2026) with over 13,000 multiple-choice questions across 11 low-resource Indic languages. In Nepali it reports Gemma-3-12B at 34.8% and Gemma-3-27B at about 39%, with the 4B model under 30%. For comparison, it gives Llama-3.3-70B 35.0%, GPT-5 43.4% and Gemini-2.5 52.0%, the best score. These are Gemma 3 results. We did not find a published Nepali benchmark for Gemma 4.

A community experiment. A developer published a small Nepali adapter for Gemma 4 E2B trained on 468 examples. On 60 English test prompts, the share of Devanagari characters in answers rose from 0.000 to 0.814. The author notes that the adapter changes language preference, not fluency, and that Nepali quality comes from the base model.

The connectivity picture. DataReportal's Digital 2026: Nepal, published November 8, 2025 with October 2025 data, reports 29.6 million people, 77% living in rural areas, 16.6 million internet users (56%), about 13.0 million offline, and 32.4 million mobile connections. The regulator's figures, reported by Nepal News, show 31.3 million broadband subscriptions in the period to August 16, 2026, but the article notes many people hold more than one connection. The Kathmandu Post reported 26.58 million 4G users in mid-March 2026, and quoted a villager saying one operator's network does not work in their village.

What the Technology Does

An open-weight model is a set of files you download and run on your own hardware. Your text goes in, the model processes it on your device or server, and an answer comes out, with no call to a cloud service. The small Gemma 4 models are sized for this: they are meant to fit on a phone or a laptop.

That changes two things for Nepal. First, privacy: documents can stay on your own machine. Second, connectivity: after the one-time download, the model can work with a weak or intermittent connection. What it does not change is the model's knowledge of Nepali, which was fixed when it was trained.

What Changed

Two years ago, running a capable model locally needed expensive hardware. Google now says its edge models are designed for phones and laptops, and the licence (Apache 2.0) allows commercial use. At the same time, researchers are starting to measure Nepali directly, as IndicParam does, and community members are already adapting the models, as the adapter above shows. Connectivity has also grown: 4G users rose from 25.11 million in March 2025 to 26.58 million in March 2026, according to the Kathmandu Post.

Who Benefits

If the language quality holds up, the people who gain most are those who are poorly served by cloud services today:

• Rural users and field workers who have a phone but an unreliable connection.
• Organisations with sensitive documents, such as clinics or schools, that prefer not to send data abroad.
• Small developers and startups who can build on open files without a per-request fee.

These are plausible benefits, not proven outcomes. We found no study measuring them in Nepal.

Limitations and Open Questions

• Nepali quality is the main gap. IndicParam scores for Gemma 3 in Nepali are well below the best model, and the benchmark covers exam-style questions drawn from linguistics and literature papers, so it may not reflect everyday use. The authors also say they used a zero-shot, log-likelihood setup that could give different results under other prompting methods.
• No Gemma 4 evidence yet. Google lists 35+ languages as supported out of the box. We could not confirm that Nepali is one of them.
• Script is not skill. The community adapter proves the model can write Devanagari, not that the answers are accurate.
• Hardware and memory. The model card gives no memory requirements, so you need to test what runs on the phones and laptops your users actually have.
• Connectivity is uneven. Most people with connections are on mobile, and coverage varies by operator and village.
• Responsibility moves to you. With an open model you run, update and secure it yourself.

What Happens Next

Watch for a published Nepali benchmark for Gemma 4, and for Nepali-focused open models. One example is Arkios, a 2026 arXiv paper describing an open English-Nepali model trained from scratch; we could not read its results, so we make no claims about its quality. For a team that wants to try this now, the practical step is a small pilot: take 50 real Nepali questions or documents from your own work, run them through an open model and a cloud model, and have a native speaker score the answers.

At Zunkiree Labs we would start from that kind of pilot before recommending either route.

The Short Version

Open models make it possible to run AI on your own device, which fits Nepal's privacy needs and uneven connectivity. But the evidence on Nepali quality is thin: the closest benchmark shows open Gemma 3 models well below the best systems, and there is no Gemma 4 Nepali result yet. Test before you build.

Read more in our State of AI in Nepal overview, our explainer on what Gemma 4 means for a business, and why AI matters for Nepal.

Related Insights

• Gemma 4: What an Apache 2.0 Open Model Means for Your Business
• Gemini 4 Argon: How to Read a Frontier Model Release
• Why AI Matters for Nepal: What the Evidence Shows

Sources

• Google, Gemma 4: Byte for byte, the most capable open models, April 2, 2026
• Google AI for Developers, Gemma 4 model card, accessed October 2026
• Maheshwari et al., IndicParam: Benchmark to evaluate LLMs on low-resource Indic languages, January 2026
• Hugging Face, saliltambe/gemma-4-E2B-it-nepali-lora, community adapter, accessed October 2026
• Regmi, Pudasaini and Pun, Arkios: An Open Bilingual English-Nepali Language Model, arXiv, 2026
• DataReportal, Digital 2026: Nepal, November 8, 2025
• Nepal News, Nepal's internet market hits 31.3 million subscriptions, August 25, 2026
• The Kathmandu Post, 4G keeps growing fast in Nepal, while users still face slow speeds and patchy coverage, April 14, 2026