Three open-weight AI models released in one week, with licensing and memory needs varying widely

Three open-weight models arrived within a single week from different labs, sized at 35B, 560B, and 816B parameters. The author argues that parameter counts alone are misleading, and that memory requirements, license terms, and the number of parameters active per token matter more for real deployment. MiniCPM-V 4.7 from OpenBMB runs on mobile hardware under Apache 2.0, while Ling-3.1-flash from inclusionAI uses an MIT license and a 1M-token context window. Meta's Muse Spark Open carries a custom license that requires attribution and a separate agreement for large-scale hosted services, showing that open weights do not automatically mean open source.