Show HN: NepaliGPT – Open-Source Nepali and English Language Model (MIT (huggingface.co)

PrinceLab’s NepaliGPT is an open-source, MIT-licensed bilingual language model finetuned from Meta’s meta-llama/Llama-3.1-8B-Instruct to support Nepali and English. Built as an instruction-tuned Llama 3.1 8B variant, it targets practical local-language applications (chatbots, education, translation, and small-scale production deployments) while keeping model size and compute needs modest. The project notes training was completed roughly 2× faster using Unsloth together with Hugging Face’s TRL library, suggesting a reproducible, cost- and time-efficient finetuning pipeline for practitioners. For the AI/ML community, NepaliGPT is significant because it provides a permissively licensed, ready-to-use foundation for building Nepali language capabilities—an underrepresented language in open model ecosystems. Key technical details: base model = Llama-3.1-8B-Instruct (instruction-tuned), finetuned with TRL+Unsloth acceleration, MIT license enables commercial and research reuse, and early uptake is small (24 downloads last month). Important implications include easier localization work, fast iteration on domain-specific finetuning, and lower barriers to deploying Nepali NLP systems; however, adopters should still validate performance, biases, safety, and benchmark the model on Nepali-specific tasks before production use.