Arcee AI

Customizing models and building new ones

Arcee AI develops language models for enterprise use. Its work spans distillation, post-training and model customization, as well as pretraining new models. Those approaches change different parts of a model's development: adapting an existing foundation is distinct from training the foundation itself. [1][2]

On September 10, 2024, Arcee launched SuperNova, a 70-billion-parameter model combining distillation from Llama 3.1 405B with instruction tuning and other post-training steps. It offered a chat interface, API access and model weights through AWS Marketplace for deployment within a customer's AWS environment. This was commercial access to weights, rather than an unrestricted public download. [1]

Extending into pretraining

In December 2025 Arcee introduced Trinity Nano and Mini, describing models pretrained from scratch. The announcement released a 6-billion-parameter Nano and a 26-billion-parameter mixture-of-experts Mini, with about 1 billion and 3 billion active parameters respectively. It discussed a future Large model separately. This illustrates why the company's history should distinguish released checkpoints from planned additions, and each model's own access terms. [2]