Many in the industry believe the winners of the AI model market have already been decided: Big Tech will own it, meaning Google, Meta, Microsoft, and a bit of Amazon, along with their model makers of choice, largely OpenAI and Anthropic. But the tiny 30-person startup Arcee AI disagrees. The company just released a truly and permanently open, Apache-licensed general-purpose foundation model called Trinity. Arcee claims that at 400 billion parameters, it is among the largest open source foundation models ever trained and released by a U.S. company.
Arcee says Trinity compares 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, according to benchmark tests conducted using base models with very little post-training. Like other state-of-the-art models, Trinity is geared for coding and multi-step processes like agents. However, despite its size, it is not a true state-of-the-art competitor yet because it currently supports only text.
More modes are in the works. A vision model is currently in development, and a speech-to-text version is on the roadmap, according to CTO Lucas Atkins. In comparison, Meta’s Llama 4 Maverick is already multi-modal, supporting text and images. But before adding more AI modes to its roster, Arcee says it wanted a base large language model that would impress its main target customers: developers and academics. The team particularly wants to woo U.S. companies of all sizes away from choosing open models from China.
The benchmarks show that the Trinity base model, currently in preview while more post-training takes place, is largely holding its own and, in some cases, slightly besting Llama on tests of coding and math, common sense, knowledge, and reasoning. The progress Arcee has made so far to become a competitive AI lab is impressive. The large Trinity model follows two previous small models released in December: the 26-billion-parameter Trinity Mini, a fully post-trained reasoning model for tasks ranging from web apps to agents, and the 6-billion-parameter Trinity Nano, an experimental model designed to push the boundaries of models that are tiny yet chatty.
The kicker is that Arcee trained them all in six months for 20 million dollars total, using 2,048 Nvidia Blackwell B300 GPUs. This came out of the roughly 50 million dollars the company has raised so far, said founder and CEO Mark McQuade. That kind of cash was a lot for the startup, said Atkins, who led the model-building effort, but he acknowledged that it pales in comparison to how much bigger labs are spending.
The six-month timeline was very calculated, said Atkins, whose career before large language models involved building voice agents for cars. He described the team as a younger startup that is extremely hungry, with a tremendous amount of talent and bright young researchers who rose to the occasion with many sleepless nights and long hours.
McQuade, previously an early employee at open source model marketplace Hugging Face, says Arcee did not start out wanting to become a new U.S. AI lab. The company was originally doing model customization for large enterprise clients like SK Telecom, taking open source models from others and post-training them for a company’s intended use. But as their client list grew, the need for their own model became a necessity, and McQuade was worried about relying on other companies. At the same time, many of the best open models were coming from China, which U.S. enterprises were leery of or barred from using.
It was a nerve-wracking decision, with McQuade noting that fewer than 20 companies in the world have ever pre-trained and released their own model at the size and level Arcee was targeting. The company started small at first with a tiny 4.5-billion-parameter model created in partnership with training company DatologyAI. The project’s success then encouraged bigger endeavors.
But if the U.S. already has Llama, why does it need another open weight model? Atkins says by choosing the open source Apache license, the startup is committed to always keeping its models open. This comes after Meta CEO Mark Zuckerberg indicated his company might not always make all of its most advanced models open source. Atkins states that Llama can be looked at as not truly open source, as it uses a Meta-controlled license with commercial and usage caveats, which has caused some open source organizations to claim that Llama is not open source compliant at all.
All Trinity models, large and small, can be downloaded for free. The largest version will be released in three flavors. Trinity Large Preview is a lightly post-trained instruct model, geared for general chat usage. Trinity Large Base is the base model without post-training. Then there is TrueBase, a model without any instruct data or post-training, so enterprises or researchers that want to customize it will not have to unroll any data, rules, or assumptions.
Arcee AI will eventually offer a hosted version of its general-release model for what it says will be competitive API pricing. That release is up to six weeks away as the startup continues to improve the model’s reasoning training. API pricing for Trinity Mini is set, and there is a rate-limited free tier available. Meanwhile, the company still sells post-training and customization options.