Sovereign stack responses
INSAIT BgGPT 3.0 (March 2026)
Claim
INSAIT released BgGPT 3.0 in three sizes — 4 billion, 12 billion, and 27 billion parameters — built on Google’s Gemma 3 and further trained on more than 100 billion words of Bulgarian text. It can handle images as well as text, remembers up to about 131,000 tokens of context at once, was pre-trained on data through May 2025, and had its instruction tuning finished through October 2025. It comes in full precision, a compressed (GPTQ W4A16) version, and the GGUF format. The earlier BgGPT (2023) was based on Gemma 2 and used about 85 billion Bulgarian and 15 billion English words, trained with the “Branch-and-Merge” method that INSAIT researchers published at the EMNLP 2024 conference. INSAIT runs its own dedicated systems for projects with Bulgaria’s National Revenue Agency and National Audit Office.
Sources
- INSAIT, “INSAIT unveils a new generation of AI,” press release, March 2026 ↗
- Bulgarian News Agency (BTA), “Bulgaria Among Few Countries with Publicly Built Sovereign AI Offered Free to Citizens,” March 2026
- INSAIT-Institute/BgGPT-Gemma-3-27B-IT model card on Hugging Face ↗
- Google DeepMind, “INSAIT creates leading Bulgarian-first LLM with Gemma 2” ↗
Referenced in
- Framework §2
- AI-SAF-N Models pillar
- AI-SAF-N Case Study (Part IV)
- Refining §2
- §5.1