Many in the industry think AI model market winner it has been decided: Big Tech will have it (Google, Meta, Microsoft, a little Amazon) along with selected model makers, mostly OpenAI and Anthropic.
But small 30-person startup Arcee AI disagrees. The company has just released a general purpose foundation model (Apache license) with a general and permanent purpose called Trinity, and Arcee claims that at 400B parameters, it is one of the largest open source foundation models ever trained and released by a US company.
Arcee says Trinity compares to Meta’s Llama 4 Maverick 400B, and Z.ai GLM-4.5, a high-performance open-source model from China’s Tsinghua University, according to benchmark tests conducted using the base model (a little post-training).
Like other state-of-the-art (SOTA) models, Trinity is geared towards coding and multi-step agent-like processes. However, despite its size, it is not a true SOTA competitor as it currently only supports text.
Another mode is in the works — a vision model is currently under development, and a speech-to-text version is on the roadmap, CTO Lucas Atkins told TechCrunch (pictured above, left). In comparison, Meta’s Llama 4 Maverick is multi-modal, supporting both text and images.
But before adding another AI mode to the roster, Arcee said, he wanted a base LLM that would impress his main target customers: developers and academics. The team especially wants to attract US companies of all sizes to avoid opting for open models from China.
“Ultimately, the winners of this game, and the only way to really win in terms of usage, is to have the best open weight model,” Atkins said. “To win the hearts and minds of developers, you have to give your best.”
Techcrunch event
San Francisco
|
13-15 October 2026
The benchmarks show that the Trinity-based model, now in preview while more post-training, generally holds its own and, in some cases, slightly outperforms Llama in tests of coding and math, common sense, knowledge and reasoning.
The progress Arcee has made so far in becoming a competitive AI Lab is impressive. The great Trinity model follows two previous small models released in December: 26B-parameter Trinity Mini, a fully post-trained reasoning model for tasks ranging from web apps to agents, and 6B-parameter Trinity Nano, an experimental model designed to push the boundaries of small but chatty models.
The kicker is, Arcee trained everything in six months for $20 million in total, using 2,048 Nvidia Blackwell B300 GPUs. This is out of approximately $50 million the company has raised so far, said founder and CEO Mark McQuade (pictured above, right).
This kind of cash is “a lot for us,” said Atkins, who led the model-building effort. Still, he admits it doesn’t compare to the larger number of labs today.
The six-month timeline is “very calculated,” says Atkins, whose career before his LLM involved building voice agents for cars. “We are a very hungry younger startup. We have a lot of talent and bright young researchers, when given the opportunity to spend money and train a model of this size, we believe they will rise to the occasion.
McQuade, previously an early employee at open-source model marketplace HuggingFace, said Arcee didn’t want to be the new US AI Lab: The company initially customized models for large corporate clients like SK Telecom.
“We only do post training. So we will take the great work of others: We will take the Llama model, we will take the Mistral model, we will take the Qwen model that is open source, and we will send training to make it better” for the company intended to use, he said, including doing reinforcement learning.
But as his client list grew, Atkins said, the need for his own model became a necessity, and McQuade worried about relying on other companies. At the same time, many of the best open models come from China, which US companies do not know about, or are prohibited from using.
It was a nerve-wracking decision. “I think there are less than 20 companies in the world that have ever trained and released their own model” in the size and level that Arcee is gunning for, McQuade said.
The company started small at first, trying its hand at a small 4.5B model created in partnership with training company DatologyAI. The success of the project then encourages larger efforts.
But if the US already has the Llama, why the need for another open weight model? Atkins said by choosing the open source Apache license, the developer is committed to keeping the model open. This happened after Meta CEO Mark Zuckerberg last year show the company can not always making all the most advanced models open source.
“Llama can be seen as non-open source because it uses a Meta-controlled license with commercial and usage warnings,” he said. This has caused some open source organizations to claim that Llama is not open source compliant at all.
“Arcee exists because the US needs to open a permanent, Apache-licensed, alternative-material border that can actually compete on the border today,” McQuade said.
All Trinity models, large and small, can be downloaded for free. The biggest version will be released in three flavors. Trinity Large Preview is a lightly trained instruction model, meaning it has been trained to follow human instructions, not just predict the next word, used for general conversational use. Trinity Large Base is a base model without post training.
Then we have TrueBase, a model that has instruction data or post training so that companies or researchers who want to manage it do not have to disclose data, rules or assumptions.
Acree AI will eventually offer a hosted version of the public release model for, say, a competitive API price. The release is up for another six weeks as the startup continues to improve the model’s reasoning training.
The API price for Trinity-Mini is $0.045 / $0.15, and there is also a limited free tier. Meanwhile, the company still sells post-training and customization options.