The open AI model landscape is becoming increasingly diverse, shifting from dominance by a few Chinese players to a broader mix of organizations including sovereign AI initiatives, Big Tech, and product companies.

  • NVIDIA released the Nemotron-3-Ultra-550B-A55B-BF16, utilizing LatentMoE for speed and adopting the OpenMDW license tailored for model weights.
  • CohereLabs released its flagship Command A+ under Apache 2.0, offering multi-modal, multi-lingual, and agentic capabilities in a 218B-A25B MoE architecture.
  • zai-org released GLM-5.2, described as genuinely usable for everyday work with minimal regression compared to the best closed models.

This diversity of makers and motivations strengthens the open ecosystem by fostering innovation through shared training methods and architectures, while restrictions on open-source development are argued to be unsafe and anti-freedom.