Pre-anneal Trinity Nano and Mini base checkpoints go live

Arcee launched its Trinity family of open-weight mixture-of-experts models on December 1, 2025, including Trinity-Mini (26 billion total parameters, 3 billion active) and a Trinity-Nano preview (6 billion total, 1 billion active). Ten days later, the company released 'pre-anneal' base checkpoints for both sizes.

Training a model typically ends with a learning-rate anneal, a final phase that gradually lowers the training step size to lock in and stabilize the model's behavior. Once a checkpoint has gone through that phase, continuing to train it further tends to be less effective, since the weights have already settled. A pre-anneal checkpoint is captured just before that step, while the weights are still comparatively fluid.

Arcee says the pre-anneal Nano and Mini checkpoints keep the same math and coding exposure as the released bases, but are meant specifically for teams planning to continue pretraining or otherwise substantially adapt the model, rather than for immediate use as-is.