Arcee releases Trinity-Large-Thinking, an open-source 399B reasoning model for enterprises

Arcee published Trinity-Large-Thinking, a 399-billion-parameter text reasoning model under the Apache 2.0 license.

The model was trained in a 33-day run on 2048 NVIDIA B300 GPUs after Arcee committed $20 million to the effort.

Trinity uses a Mixture-of-Experts design with only 1.56% of parameters active per token and includes SMEBU for expert stability.

Arcee trained on 20 trillion tokens split between curated web data and synthetic data from DatologyAI to improve reasoning.