Trinity Mini

Trinity Large Preview is a 400B-parameter (about 13B active) frontier Trinity MoE model with long-context comprehension, tuned for strong reasoning, coding and multi-step agents, served via hosted preview APIs.

Trinity Nano Base is the base 6B (1B active) Trinity Nano checkpoint, pre–fine-tuning, meant to be domain-tuned rather than used directly for chatting, trained on about 10T tokens under Apache-2.0.

Jamba2 is AI21’s open-source enterprise LLM family, optimized for reliability, steerability, and grounding, with compact 3B and 52B variants and long context for production workflows.

The largest Ministral 3 model offers frontier text and vision capabilities comparable to larger 24B models. Edge-optimized for single GPU deployment (24GB VRAM in FP8), it delivers state-of-the-art performance for chat, document analysis, and complex reasoning tasks with multilingual support across 40+ languages.

Cosmos Reason is NVIDIA’s open, customizable 7B-parameter reasoning VLM for physical AI and robotics. It learns space/time/physics “common sense,” plans actions from video/images, and was trained with SFT + RL. Available under the NVIDIA Open Model License on GitHub/Hugging Face and as an NVIDIA NIM.