Soro: A Lightweight Foundation Model and Chatbot for Tajik

Researchers have developed Soro, a family of Tajik-specialized conversational AI models built from Gemma 3 checkpoints and trained on a 1.9-billion-token Tajik corpus. The models outperform same-size baselines on new Tajik benchmarks covering general knowledge and exam domains, while retaining English performance. Soro's quantized versions enable deployment on low-resource devices, supporting an ongoing education pilot in Tajikistan.

arXiv:2605.27379v1 Announce Type: new Abstract: We present Soro, a family of Tajik-specialized conversational large language models LLMs designed for real-world deployment under tight compute and connectivity constraints in Tajikistan. Starting from open-weight Gemma 3 checkpoints, we perform Tajik-only continual pretraining on a curated 1.9-billion-token corpus spanning filtered web text, PDF documents, and curriculum-aligned educational materials, followed by supervised instruction tuning on 40K Tajik teacher-style examples. To enable rigorous evaluation despite the limited coverage of Tajik in standard benchmarks, we introduce a suite of Tajik benchmarks covering general knowledge, linguistic competence, and school- and university entrance-exam domains, and we open-source them on Hugging Face. Across these Tajik benchmarks, Soro substantially outperforms same-size Gemma 3 baselines while retaining strong English performance on standard datasets. We further show that FP8 and INT4 quantization of Soro preserves most Tajik-language gains while reducing memory requirements for edge deployment, supporting an ongoing education-sector pilot and planned scale-out across schools in Tajikistan.