Read more: savetibet.org/deepzang-the...
Read more: savetibet.org/deepzang-the...
open.substack.com/pub/hrichina...
China's claiming it's the "world's first" Tibetan #LLM.
It's not. Here's the technical and political reality. 🧵
web.archive.org/web/20260316...
China's claiming it's the "world's first" Tibetan #LLM.
It's not. Here's the technical and political reality. 🧵
web.archive.org/web/20260316...
stratnewsglobal.com/asia/in-tibe...
stratnewsglobal.com/asia/in-tibe...
www.tibetanreview.net/china-launch...
www.tibetanreview.net/china-launch...
#DeepZang inherits DeepSeek's architecture and its CAC-mandated content restrictions.
Different tech stacks. Different constraints.
#DeepZang inherits DeepSeek's architecture and its CAC-mandated content restrictions.
Different tech stacks. Different constraints.
#DeepZang was trained on ~70m Tibetan-Chinese parallel texts + 30,500 hours of #Tibetan speech across Ü-Tsang, Kham & Amdo dialects. web.archive.org/web/20260511...
#DeepZang was trained on ~70m Tibetan-Chinese parallel texts + 30,500 hours of #Tibetan speech across Ü-Tsang, Kham & Amdo dialects. web.archive.org/web/20260511...
www.tibetanreview.net/china-launch...
www.tibetanreview.net/china-launch...
Timeline:
Nov 2024: Monlam Melong (exile, Dharamsala) — open source, on #HuggingFace
Nov 2025: 阳光清言 (Tibet University / Zhipu AI) — 100B params
Mar 2026: #DeepZang
Timeline:
Nov 2024: Monlam Melong (exile, Dharamsala) — open source, on #HuggingFace
Nov 2025: 阳光清言 (Tibet University / Zhipu AI) — 100B params
Mar 2026: #DeepZang
Translation: fine-tuned #DeepSeek with a 70M Tibetan-Chinese parallel corpus + 30,500 hours of dialect speech data.
Translation: fine-tuned #DeepSeek with a 70M Tibetan-Chinese parallel corpus + 30,500 hours of dialect speech data.
www.tibetanreview.net/chinas-deepz...
www.tibetanreview.net/chinas-deepz...
www.rfa.org/tibetan/sarg...
www.rfa.org/tibetan/sarg...