Trinity Large Preview
Arcee AI · 2026-01-27 · 398.6B parameters
Trinity Large Preview is Arcee AI's lightly post-trained, chat-ready sparse mixture-of-experts checkpoint with approximately 13B active parameters per token. The underlying Trinity Large model was reported as pretrained on 17T tokens and then context-extended at a 256K training length for documented 512K-token inference; this Preview checkpoint was instruction-tuned on a reported 20B tokens. Arcee explicitly released it as non-reasoning, and its launch API served it at 128K context with 8-bit quantization.