Arcee AI, a 30-person startup, has released Trinity, a general purpose foundation model with 400B parameters that it describes as a truly and permanently open model under an Apache license. The company says Trinity compares to Meta’s Llama 4 Maverick 400B and Z.ai’s GLM-4.5 from Tsinghua University, citing benchmark tests on base models with very little post-training. Trinity is currently text only, but Arcee is positioning it as a coding capable, agent friendly model designed to appeal primarily to developers and academics, and to offer U.S. organizations an alternative to open models originating in China.
The startup has developed a small portfolio of Trinity models, starting with a tiny 4.5B model built with DatologyAI before moving to larger systems. In December, it released the 26B parameter Trinity Mini, which is a fully post-trained reasoning model for tasks ranging from web apps to agents, and the 6B parameter Trinity Nano, an experimental model meant to explore how small models can still be conversational. Arcee says all of these models, including the new 400B version, were trained in six months for 20 million total, using 2,048 Nvidia Blackwell B300 GPUs, out of roughly 50 million the company has raised so far.
Benchmarks cited by Arcee indicate that the Trinity base model, which remains in preview while further post-training is underway, is largely competitive with and in some cases slightly outperforms Llama on coding, math, common sense, knowledge, and reasoning tests. The largest Trinity model will ship in three variants: Trinity Large Preview, a lightly post-trained instruct model optimized for general chat; Trinity Large Base, the unpost-trained base model; and TrueBase, a version with no instruct data or post training so enterprises and researchers can customize it without having to undo prior alignment or assumptions. All Trinity models can be downloaded for free, and Arcee plans to offer a hosted general release model with what it describes as competitive API pricing within up to six weeks as it continues improving reasoning training.
Arcee’s founders argue that the United States needs a permanently open, Apache licensed, frontier grade model that is not subject to the shifting licensing strategies of large platforms. They contrast their approach with Meta’s Llama, which they characterize as using a Meta controlled license with commercial and usage restrictions, and point to concerns in some U.S. enterprises about relying on leading open models produced in China. The company, which previously focused on post-training and customization for large enterprises using models like Llama, Mistral, and Qwen, decided to build its own models to reduce dependence on upstream providers and to better address customer needs. Arcee will continue to sell post-training and customization services alongside its new models and currently lists API pricing for Trinity Mini as 0.045 / 0.15, with a rate-limited free tier also available.
