Arcee AI, the American lab behind the open Trinity models

From post-training to in-house models

Arcee AI is a San Francisco-based lab that describes itself as a US “open intelligence lab.” Its official pages do not state when it was founded. The company sums up its own path: it began in post-training, meaning the fine-tuning of models built by others, before moving upstream to pre-train its own. Its core belief is that progress will be defined not just by size, but by efficiency, adaptability and the ability to own what you build.

In September 2026, Arcee announced a Series B round to fund its next generations of models and its work with the US Department of Energy.

The Trinity family and other models

As of October 22, 2026, the main lineup is called Trinity:

• Trinity Large Thinking, the most capable reasoning model, geared toward agents;
• Trinity Mini, a balance of performance and speed;
• Trinity Nano, a small language model light enough to run on a device.

Arcee also launched AFM-4.5B in July 2025 and Genesis-Science-1, an open model designed for scientific research. The company emphasizes reliable tool calling and structured outputs, two key requirements for building agents.

A firmly open policy

All flagship models are released as open weights. In May 2026, Trinity moved to OpenMDW-1.1, a permissive Linux Foundation license built to cover an entire model release (weights, code, configurations, data) in one document. Arcee presents this openness as a way to avoid vendor lock-in. That positioning sets it apart: it presents itself as one of the few US players releasing large models under a permissive license.

How to access the models

• Arcee’s free web chat, for a first try.
• The API, including the Open Models API in beta since August 2026, documented at docs.arcee.ai.
• Hugging Face, to download the weights and host them yourself.

To compare Trinity with other open models, see our page on open-source models.