Large language models (LLMs) have revolutionized how we interact with technology, but their prowess has been largely confined to English. What about the millions of people who speak other languages? Researchers at INSAIT are tackling this challenge head-on with BgGPT, a project dedicated to extending the capabilities of powerful LLMs to Bulgarian. They’re not just translating English LLMs—they’re building a model that truly *understands* and *generates* high-quality Bulgarian text while retaining and even improving the original English capabilities. This is a significant hurdle, as adapting existing LLMs to new languages often leads to a decline in performance in the original language, a phenomenon known as catastrophic forgetting. To combat this, the team is using innovative techniques like Branch-and-Merge, a continual learning strategy that minimizes performance loss while maximizing gains in the new language. They've also curated a massive dataset of over 100 billion tokens of Bulgarian and English text. BgGPT isn’t just a research project—it’s a real-world application. The models power a Bulgarian chat service, making powerful AI accessible to users without specialized hardware. The team has also focused on educational applications, benchmarking BgGPT against state-of-the-art models using exam questions provided by the Bulgarian Ministry of Education. Results are impressive. BgGPT outperforms larger multilingual models like Qwen-2.5 and Llama-3.1 in several key benchmarks, demonstrating its specialized proficiency in Bulgarian. While the primary focus is Bulgarian, the researchers believe their methods can be adapted for other lower-resource languages, opening doors for wider access to powerful AI tools. This is more than just language translation; it's about bridging the digital divide and bringing the transformative power of AI to everyone, regardless of their language.