ModelBest has open-sourced its first Mixture-of-Experts (MoE) multimodal model, MiniCPM-V 4.7. This model is available in two sizes, 35B-A3B and 1B. The larger MoE version can reportedly run on a smartphone with approximately 3GB of memory, achieving decoding speeds of over 13 tokens per second. AI
IMPACT Enables on-device multimodal AI capabilities with low memory requirements.
RANK_REASON Open-source release of a new model by a non-frontier lab. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →