Trinity Large Preview is Arcee AI's flagship open-weights release, a roughly 398-billion-parameter sparse Mixture-of-Experts model that activates about 13 billion parameters per token. It was trained on more than 17 trillion tokens, making it the largest member of the Trinity family and the foundation for several downstream variants. The Preview checkpoint is a lightly post-trained, chat-ready build sitting on top of Trinity-Large-Base, and it is described as still undergoing active reinforcement-learning refinement. A companion technical report on GitHub walks through the full training recipe for anyone who wants to study the lineage behind this checkpoint.
The same training run produced two sister checkpoints worth knowing about: Trinity-Large-Thinking, which adds reasoning-optimized, agentic post-training with extended chain-of-thought, and Trinity-Large-TrueBase, a 10-trillion-token pre-anneal snapshot that preserves the raw pretraining behavior. Practically, the Preview variant is positioned as a chat-friendly entry point that retains the family's strong long-context comprehension while keeping the parameter cost per token low, which is attractive for teams that want a large open model without paying the full inference bill. It fits well for experimentation, evaluation against other open frontier models, and workloads that benefit from a chat-ready interface backed by an openly published base.
