YuanLab Open-Sources Yuan3.0 Ultra, Joins Top 3 Trillion-Parameter Open Multimodal Models Globally
YuanLab.ai open-sourced Yuan3.0 Ultra, one of only three trillion-parameter open multimodal foundation models worldwide. It uses a unified multimodal architecture with a 103-layer MoE language backbone, optimized to 1.01T parameters (68.8B active). Its Localized Filtering Attention (LFA) boosts semantic modeling and cuts pre-training compute by 49%. It supports text, image, audio, and video I/O for content creation, intelligent interaction, and research.
Keywords：Yuan3.0 Ultra, open-source, trillion parameters, multimodal LLM, MoE architecture
new
OpenClaw may be the most important software release in history
Google Gemini 3 User Base Surges, Challenging OpenAI's Dominance
Claude Free Users Surge 60%+; Anthropic Grows Despite Pentagon Ban Risks
YuanLab Open-Sources Yuan3.0 Ultra, Joins Top 3 Trillion-Parameter Open Multimodal Models Globally
OpenAI Launches GPT-5.4 with Native Computer Control, Deep Integration into Excel & Google Sheets
Tesla Unveils Optimus Gen-3 Humanoid Robot, Announces Mass Production Plan
Google DeepMind Unveils Math Reasoning Agent Aletheia, Solving World-Class Problems
Edge AI Explodes, Honor Launches Magic8 Pro Redefining Mobile Imaging
Open-Source LLM Architecture Boom: In-Depth Review of 10 New 2026 Models
Mistral Releases Local Voice Model: Privacy-Focused, Low-Latency Real-Time Conversation