← All models

Trinity Nano Preview

Arcee AI · 2025-12-01 · 6.1B parameters

Trinity Nano Preview is Arcee AI's experimental chat-tuned sparse mixture-of-experts checkpoint with approximately 1B active parameters (about 800M non-embedding) across 56 layers, routing each token to 8 of 128 experts plus one shared expert. Arcee describes it as a personality-forward, non-thinking, download-only release that may be unstable; it was trained on a reported 10T tokens and at a 256K sequence length for 128K inference. It was announced under Apache 2.0 on December 1, 2025, and the repository adopted OpenMDW-1.1 in May 2026.

Benchmark scores