Arcee AI releases Trinity, a 400B-parameter Apache-licensed model

Arcee AI, a 30-person startup, has released a general-purpose foundation model—defined as a large-scale AI model—called Trinity under an Apache license, which is a permissive free software license. The company claims that the 400B-parameter model is among the largest open source foundation models ever trained and released by a U.S. company.

The launch represents a challenge to the AI model market, which is largely populated by Big Tech participants like Google, Microsoft, and Amazon, alongside model makers OpenAI and Anthropic. Arcee attributes Trinity as comparable to Meta’s Llama 4 Maverick 400B and Z.ai’s GLM-4.5, a model created by Tsinghua University. The company is pitching the model as a permanently open alternative to Llama, criticizing Meta’s licensing terms. Lucas Atkins, the CTO of Arcee AI, stated: “Llama can be looked at as not truly open source as it uses a Meta-controlled license with commercial and usage caveats.” Founder and CEO Mark McQuade, a former employee of Hugging Face, asserted that the startup exists to provide a permanently open, Apache-licensed alternative for the U.S. that can compete at the frontier.

The startup completed the training run using a lean team. The technical and financial specifications of the project include:

• Timeline: six months of training
• Total cost: $20 million total, out of roughly $50 million total raised by the company
• Hardware: 2,048 Nvidia Blackwell B300 GPUs

Atkins attributed the achievement to a younger, hungry team of researchers who worked long hours to train a model of this size. McQuade noted that less than 20 companies in the world have ever pre-trained and released their own model of this scale. Previously, the company focused on post-training other open models for enterprise clients like SK Telecom. It had also built a 4.5B model in partnership with training partner DatologyAI, before deciding to build its own base model to avoid relying on other companies.

Although Trinity is not a true SOTA competitor yet because it currently supports only text, Arcee plans to expand its capabilities. The company previously released the 26B-parameter Trinity Mini and the 6B-parameter Trinity Nano in December. While the models are available for download, Arcee plans to offer a version of its general-purpose model with API pricing. For context, the API pricing for Trinity Mini is $0.045 / $0.15.

Why it matters

Arcee AI is positioning its 400B-parameter model, Trinity, as a U.S.-based, permanently open-source alternative to Big Tech models and models from China, aiming to win over developers and academics.