Small startup Arcee AI built a 400B open source LLM from scratch to better meta’s Llama

Many in the industry Consider the winners in the AI ​​modeling market The decision has already been made: the big tech companies (Google, Meta, Microsoft, a little Amazon) will own it along with their preferred model makers, largely OpenAI and Anthropic.

But Arcee AI, a small 30-person startup, disagrees. The company has just released a truly open, perpetual, general-purpose (Apache License) foundation model called Trinity, and Arcee claims that, at 400B parameters, it is among the largest open source foundation models ever trained and released by a US company.

Arcee says the 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, according to benchmark tests performed with basic models (very little post-training).

Like other SOTA models, Trinity is designed for multi-step programming and agent-like operations. However, despite its size, it is not a real competitor to SOTA yet as it currently only supports text.

There are more modes in the works — a vision model is currently in development, and a speech-to-text version is on the roadmap, CTO Lucas Atkins told TechCrunch (pictured above, left). By comparison, Meta’s Llama 4 Maverick is truly multimedia, supporting text and images.

But before adding more AI positions to her list, Arcee says, she wanted a basic LLM certification that would impress her main target clients: developers and academics. The team particularly wants to lure US companies of all sizes away from opting for open models from China.

“At the end of the day, the winners 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.”

The benchmarks show that the basic Trinity model, which is currently in preview while more training is done after it, largely holds its own and, in some cases, slightly outperforms Llama on tests of coding, mathematics, common sense, knowledge, and reasoning.

The progress Arcee has made so far in becoming a competitive AI laboratory is impressive. It follows the model of the Great Trinity The previous two small models Released in December: Trinity Mini with parameter 26B, a complete post-training reasoning model for tasks ranging from web applications to agents, and Trinity Nano with parameter 6B, an experimental model designed to push the boundaries of small but talkative models.

The kicker is that Arcee trained them all up in six months for a total of $20 million, using 2,048 Nvidia Blackwell B300 GPUs. This is among the nearly $50 million the company has raised to date, said founder and CEO Mark McQuaid (pictured above, right).

That kind of money was “a lot for us,” said Atkins, who led the effort to build the model. However, he acknowledged that this amount pales in comparison to the amount major laboratories currently spend.

The six-month timeline “was very calculated,” said Atkins, whose pre-MBA career involved building audio proxies for cars. “We are a very hungry young startup. We have a tremendous amount of talent and bright young researchers who, when given the opportunity to spend this amount of money and train a model of this scale, we were confident they would rise to the occasion. And they certainly did, with many sleepless nights, and many long hours.”

McQuade, who was previously an early employee at open source prototyping market HuggingFace, says Arcee didn’t start out wanting to become a new American AI lab: The company was originally customizing prototyping for large enterprise customers like SK Telecom.

“We were just doing post-training. So we were taking the great work of others: we were taking the Llama model, we were taking the Mistral model, we were taking the Cowen model that was open source, and we would then train it to make it better” for the company’s intended use, including doing reinforcement learning.

But as their client list grew, Atkins said, the need for their own model became imperative, and McQuaid was concerned about relying on other companies. Meanwhile, many of the best open models were coming from China, which American companies were wary of, or banned from using.

It was a nerve-wracking decision. “I think there are less than 20 companies in the world that have pre-trained and deployed their own model” at the scale and scale Arcee was seeking, McQuaid said.

The company initially started small, piloting a 4.5B prototype created in partnership with training company DatologyAI. Then the success of the project encouraged greater efforts.

But if the US already has the Llama, why does it need another open-weight model? By choosing the open source Apache license, the startup is committed to always keeping its models open, Atkins says. This follows Meta CEO Mark Zuckerberg last year He noted that his company may not always be like that Make all of its most advanced models open source.

“Llama can be viewed as not being truly open source because it uses a Meta-controlled license with commercial and usage warnings,” he says. This has caused Some organizations claim open source Llama is not compatible with open source at all.

“Arcee exists because the United States needs a permanently open, licensed, border-grade Apache alternative that can actually compete on today’s borders,” McQuaid said.

All Trinity templates, large and small, are free to download. The larger 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, which prepares it for general chat use. Trinity Large Base is the basic model with no subsequent training.

Then we have TrueBase, which is a model that contains any post instruction or training data, so organizations or researchers who want to customize it won’t have to expose any data, rules, or assumptions.

Acree AI will eventually offer a hosted version of its public release model for competitive API pricing. This release will take up to six weeks as the startup continues to improve inference training of the model.

API pricing for Trinity-Mini is $0.045 / $0.15, and a limited-priced free tier is also available. Meanwhile, the company still sells post-training and customization options.