A small, 30-person startup named Arcee AI is challenging the prevailing industry belief that the artificial intelligence model market will be dominated solely by Big Tech giants like Google, Meta, Microsoft, and Amazon, along with their preferred model makers such as OpenAI and Anthropic. Arcee AI has recently unveiled Trinity, a general-purpose, foundation model that is truly and permanently open, licensed under the Apache license.

Arcee AI claims that Trinity, at 400 billion parameters, stands as one of the largest open-source foundation models ever trained and released by a U.S. company. Benchmark tests conducted using base models (with minimal post-training) indicate that Trinity compares favorably to Meta’s Llama 4 Maverick 400B and Z.ai’s GLM-4.5, a high-performing open-source model from China’s Tsinghua University.

Like other state-of-the-art models, Trinity is designed for sophisticated tasks such as coding and multi-step processes for agents. However, despite its substantial size, it currently supports only text. Arcee AI's CTO, Lucas Atkins, has confirmed that a vision model is actively in development, and a speech-to-text version is planned for the future, aiming to expand Trinity's modalities. In contrast, Meta’s Llama 4 Maverick is already multi-modal, handling both text and images.

Arcee's strategic decision was to first establish an impressive base Large Language Model (LLM) to attract its primary target customers: developers and academics. The company specifically aims to persuade U.S. companies of all sizes to choose its models over open models originating from China. Atkins emphasized Arcee's core belief, stating, “Ultimately, the winners of this game, and the only way to really win over the usage, is to have the best open-weight model. To win the hearts and minds of developers, you have to give them the best.”

The benchmark results show that the Trinity base model, currently in a preview phase while undergoing further post-training, generally holds its own and, in some tests, slightly outperforms Llama on coding, math, common sense, knowledge, and reasoning evaluations. This progress showcases Arcee AI's significant strides toward becoming a competitive AI Lab.

The large Trinity model follows the release of two smaller models in December: the 26B-parameter Trinity Mini, a fully post-trained reasoning model suitable for various tasks from web applications to agents, and the 6B-parameter Trinity Nano, an experimental model designed to explore the capabilities of tiny yet conversational models.

Notably, Arcee AI accomplished the training of all these models within six months, with a total cost of $20 million, utilizing 2,048 Nvidia Blackwell B300 GPUs. This expenditure comes from approximately $50 million the company has raised to date, as stated by founder and CEO Mark McQuade. Atkins, who led the model-building effort, acknowledged that while $20 million was a significant investment for Arcee, it is modest compared to the spending of larger labs. The six-month timeline was meticulously planned, according to Atkins, who attributed the swift success to a highly motivated team of talented young researchers who rose to the occasion through many sleepless nights and long hours.

Mark McQuade, a former early employee at the open-source model marketplace Hugging Face, revealed that Arcee AI did not initially intend to become a new U.S. AI lab. The company originally specialized in model customization, providing post-training services for large enterprise clients like SK Telecom. This involved taking existing open-source models such as Llama, Mistral, and Qwen, and enhancing them for specific client uses, including reinforcement learning.

However, as Arcee's client base expanded, the necessity for their own proprietary model became evident. McQuade grew concerned about the risks of relying on other companies' models, especially as many of the best open models were originating from China, which U.S. enterprises were hesitant or even prohibited from using. The decision to pre-train their own model was a daunting one, as McQuade noted that fewer than 20 companies globally had successfully pre-trained and released a model at the scale Arcee was targeting. The journey began with a smaller, 4.5B model, created in partnership with training company DatologyAI, whose success encouraged their larger endeavors.

Addressing the question of why another open-weight model is needed given Llama's presence, Atkins explained that Arcee's choice of the Apache license signifies a commitment to permanent openness for its models. This contrasts with Meta CEO Mark Zuckerberg's previous indications that not all of Meta's most advanced models might remain open source. Atkins argues that Llama may not be considered truly open source due to its Meta-controlled license, which includes commercial and usage caveats, leading some open-source organizations to question its compliance. McQuade affirmed, “Arcee exists because the U.S. needs a permanently open, Apache-licensed, frontier-grade alternative that can actually compete at today’s frontier.”

All Trinity models, both large and small versions, are available for free download. The largest version of Trinity will be released in three distinct configurations: Trinity Large Preview, a lightly post-trained instruct model designed to follow human instructions for general chat usage; Trinity Large Base, which is the foundational model without any post-training; and TrueBase, a model completely devoid of instruct data or post-training, ideal for enterprises or researchers who wish to customize it without having to undo existing data, rules, or assumptions.

Arcee AI plans to eventually offer a hosted version of its general-release model with competitive API pricing. This general release is anticipated within six weeks, as the startup continues to refine the model's reasoning training. Currently, API pricing for Trinity Mini is $0.045 / $0.15, and a rate-limited free tier is also available. Furthermore, the company continues to provide its core services of post-training and customization.