Arcee AI Releases Trinity, a 400B-Parameter Open-Source LLM

Arcee AI Introduces Trinity

Arcee AI, a small U.S. startup with roughly thirty employees, announced the release of Trinity, a 400‑billion‑parameter large language model (LLM). The model is offered under a permissive Apache license, making it permanently open source. According to the company, Trinity competes with Meta’s Llama 4 Maverick 400B and the Chinese‑origin GLM‑4.5 in benchmark evaluations that focus on base model performance with minimal post‑training.

Performance and Capabilities

Early benchmark results indicate that Trinity’s base model holds its own against leading models in several domains, including coding, mathematics, common‑sense reasoning, factual knowledge, and logical reasoning. At present, Trinity supports only text input and output, but the roadmap includes a vision model and a speech‑to‑text variant. This contrasts with Meta’s Llama 4 Maverick, which already supports both text and images.

Target Audience and Market Position

The startup’s primary audience consists of developers and academic researchers. Arcee AI aims to provide a high‑quality, open‑weight alternative to models that are either proprietary or subject to licensing restrictions. The company emphasizes that a permanently open Apache‑licensed model can win the “hearts and minds” of U.S. developers who are hesitant to rely on open models from China.

Model Variants and Availability

Trinity will be released in three distinct versions. “Trinity Large Preview” is a lightly post‑trained instruction‑following model intended for general chat use. “Trinity Large Base” is the pure base model without any post‑training. “TrueBase” removes all instruction data and post‑training, allowing enterprises and researchers to customize the model without inheriting any pre‑imposed rules. All versions can be downloaded for free, and a hosted API offering with competitive pricing is expected to launch within six weeks.

Training Infrastructure and Cost

The model was trained over a six‑month period using 2,048 Nvidia Blackwell B300 GPUs. The total expense for training Trinity was $20 million, funded from the roughly $50 million the company has raised to date. The training effort was led by CTO Lucas Atkins, who previously built voice agents for automotive applications.

Company Background and Strategy

Founder and CEO Mark McQuade, a former early employee of Hugging Face, explained that Arcee AI originally focused on post‑training and customization services for large enterprise clients such as SK Telecom. The company would take existing open‑source models—Llama, Mistral, Qwen—and improve them through reinforcement learning and other techniques. As the client base expanded, the need for a proprietary model grew, prompting the decision to develop Trinity.

Open‑Source Philosophy

Arcee AI chose the Apache license to ensure that Trinity remains fully open source, differentiating it from models like Meta’s Llama, which uses a Meta‑controlled license with commercial and usage caveats. The startup argues that a truly open, frontier‑grade model is essential for U.S. enterprises that are wary of relying on Chinese‑origin models.

Future Outlook

Beyond Trinity, Arcee AI plans to continue expanding its model suite with multimodal capabilities and to offer hosted API services at rates competitive with major cloud providers. The company also maintains a business line for post‑training and customization services, leveraging its expertise in tailoring models for specific enterprise applications.