Many in the industry believe that the winner of the AI model market has already been determined. Big tech will own this market (Google, Meta, Microsoft, parts of Amazon) and the model makers of their choice (mostly OpenAI and Anthropic).
But Arcee AI, a small startup with 30 employees, disagrees. The company just released a truly perpetually open (Apache License) general purpose foundation model called Trinity, which Arcee claims is one of the largest open source foundation models ever trained and released by a US company, with 400B parameters.
According to Arcee, benchmark tests conducted using the base model (with very little post-training) show that Trinity is comparable to Meta’s Llama 4 Maverick 400B and Z.ai’s GLM-4.5, a high-performance open source model from China’s Tsinghua University.
Like other state-of-the-art (SOTA) models, Trinity supports multi-step processes such as coding and agents. Still, despite its size, it’s still not a true SOTA competitor, as it only supports text at the moment.
More modes are in development. The vision model is currently in development, and a speech-to-text version is also on the roadmap, CTO Lucas Atkins told TechCrunch (pictured above, left). In comparison, Meta’s Llama 4 Maverick is already multimodal and supports text and images.
But before adding more AI modes to its roster, Arcee says it wanted a basic LLM that would impress its primary target customers: developers and academics. The team specifically wants to ensure that U.S. companies of all sizes don’t choose China’s open model.
“Ultimately, the winner in this game, and the only way to really win in usage, is to have the best openweight model,” Atkins said. “To win the hearts and minds of developers, you have to give them the best.”
tech crunch event
san francisco
|
October 13-15, 2026
Benchmarks show that the Trinity base model, currently in preview and undergoing further post-training, holds its own, and in some cases slightly outperforms Llama, on coding and math, common sense, knowledge, and reasoning tests.
The progress Arcee has made to become a competitive AI Lab is impressive. The larger Trinity model follows two previous smaller models released in December. The 26B-parameter Trinity Mini is a fully post-trained inference model for tasks ranging from web apps to agents, and the 6B-parameter Trinity Nano is an experimental model designed to push the limits of small but talkative models.
Amazingly, Arcee trained everyone in six months using 2,048 Nvidia Blackwell B300 GPUs at a total cost of $20 million. Founder and CEO Mark McQuaid (pictured above, right) says this is part of the roughly $50 million the company has raised to date.
That kind of cash “was a lot for us,” said Atkins, who led the model-building effort. Still, he acknowledged that it pales in comparison to what large research institutions currently spend.
The six-month schedule was “very calculated,” Atkins said. Prior to LLM, Atkins worked building voice agents for automobiles. “We’re a very hungry young startup. We have tremendous talent and bright young researchers. We believed that if they were given the opportunity to spend this amount of money and train a model of this scale, they would rise to the occasion. And they certainly did, even though they had sleepless nights and put in long hours.”
McQuade, who was previously an early employee at open source modeling marketplace Hugging Face, says Arcee didn’t start out with the goal of becoming America’s new AI lab. The company originally customized models for large corporate clients like SK Telecom.
“We were only doing post-training, so we took the great work of others. We took the Llama model, we took the Mistral model, we took the open source Kwen model and we post-trained it to improve it,” he said, including doing reinforcement learning for the company’s intended use.
But as the customer list grew, Atkins said, the need for a proprietary model grew and McQuaid was concerned about becoming dependent on other companies. At the same time, many of the best open models came from China, and U.S. companies were wary of them or were banned from using them.
It was a nerve-wracking decision. At the scale and level that Arcee was aiming for, “I think there are less than 20 companies in the world that have ever pre-trained and released their own models,” McQuade said.
The company started small, taking on 4.5 billion small models created in partnership with training company DatologyAI. The success of the project encouraged even larger efforts.
But if the U.S. already has Llamas, why do we need another indiscriminate weight model? Atkins said by opting for the open source Apache license, the startup is committed to keeping the model open. This comes after Meta CEO Mark Zuckerberg suggested last year that the company might not necessarily open source all of its latest models.
“Llama can be considered not truly open source because it uses a meta-managed license with commercial and usage caveats,” he says. This has led some open source organizations to claim that Llama is not open source compliant at all.
“Arcee exists because America needs a permanently open, Apache-licensed, frontier-grade alternative that can actually compete on today’s frontier,” McQuade said.
All Trinity models, large and small, are available for free download. The maximum version will be released in 3 flavors. Trinity Large Preview is a lightly post-trained instructional model. This means that it is trained to not only predict the next word, but also to follow human instructions, making it suitable for general chat use. Trinity Large Base is a base model with no post training.
Additionally, TrueBase provides arbitrary instructional data and post-training models, so companies and researchers who want to customize do not need to deploy data, rules, or assumptions.
Arcee AI says it plans to eventually offer a hosted version of its general release model at competitive API prices. Its release is up to six weeks away as the startup continues to improve the model’s inference training.
Trinity Mini has an API price of $0.045 / $0.15, with a rate-limited free tier available. Meanwhile, the company continues to sell post-training and customization options as well.
