← All models

Trinity Mini

Arcee AI · 2025-12-01 · 26.1B parameters

Trinity Mini is Arcee AI's fully post-trained reasoning model, a sparse mixture-of-experts checkpoint with approximately 3B active parameters that routes each token to 8 of 128 experts plus one shared expert. Its base model was trained on a reported 10T tokens in three phases and at a 128K sequence length, and its chat template always opens a <think> block. Arcee's December 2025 launch post announced it under Apache 2.0; the Hugging Face repository adopted OpenMDW-1.1 in May 2026.

Benchmark scores