Arcee AI's Trinity 400B debut
Tiny 30-person startup Arcee AI released Trinity, a 400B-parameter open foundation model Apache-licensed, claiming it rivals leading open models. Arcee claims Trinity is among the largest open source foundation models ever trained and released by a U.S. company. But the industry consensus had long favored Big Tech incumbents: Google, Meta, Microsoft, and their model makers OpenAI and Anthropic. Arcee disagrees.
Benchmark comparisons
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. The Trinity base model currently in preview holds its own and slightly bests Llama on tests of coding, math, common sense, knowledge, and reasoning. Despite its size, Trinity is not a true SOTA 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, CTO Lucas Atkins told TechCrunch. In comparison, Meta's Llama 4 Maverick is already multi-modal, supporting text and images. But before adding more AI modes, Arcee wanted a base LLM 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.
Training economics
Arcee trained Trinity in six months for $20 million total, using 2,048 Nvidia Blackwell B300 GPUs. This out of the roughly $50 million the company has raised so far, said founder and CEO Mark McQuade. 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 how much bigger labs are spending right now. The six-month timeline "was very calculated," said Atkins, whose career before LLMs involved building voice agents for cars. "We are a younger startup that's extremely hungry. 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 size, we trusted that they'd rise to the occasion. And they certainly did, with many sleepless nights, many long hours."
Trinity model lineup
Trinity comes in three flavors. Trinity Large Preview is a lightly post-trained instruct model, meaning it's been trained to follow human instructions, not just predict the next word, which gears it for general chat usage. Trinity Large Base is the base model without post-training. Then we have TrueBase, a model with any instruct data or post training so enterprises or researchers that want to customize it won't have to unroll any data, rules, or assumptions. All Trinity models, large and small, can be downloaded for free. The largest version will be released in these three flavors.
Product roadmap and API pricing
Arcee AI will eventually offer a hosted version of its general-release model for, it says, 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 $0.045 / $0.15, and there is a rate-limited free tier available, too. Meanwhile, the company still sells post-training and customization options. Previously Arcee was doing model customization for enterprise clients like SK Telecom. The company was originally doing post-training: taking a Llama model, a Mistral model, a Qwen model that was open source, and post-training it to make it better for a company's intended use, including doing the reinforcement learning. But as their client list grew, the need for their own model was becoming 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 were barred from using.
From customization to own model
It was a nerve-wracking decision. "I think there's less than 20 companies in the world that have ever pre-trained and released their own model" at the size and level that Arcee was gunning for, McQuade said. The company started small at first, trying its hand at a 4.5B 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. "Llama can be looked at as not truly open source as it uses a Meta-controlled license with commercial and usage caveats," he says. This has caused some open source organizations to claim that Llama isn't open source compliant at all. "Arcee exists because the U.S. needs a permanently open, Apache-licensed, frontier-grade alternative that can actually compete at today's frontier," McQuade said.
Smaller predecessors
Arcee started small at first, trying its hand at a tiny 4.5B 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. "Llama can be looked at as not truly open source as it uses a Meta-controlled license with commercial and usage caveats," he says. This has caused some open source organizations to claim that Llama isn't open source compliant at all. "Arcee exists because the U.S. needs a permanently open, Apache-licensed, frontier-grade alternative that can actually compete at today's frontier," McQuade said.
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, meaning it's been trained to follow human instructions, not just predict the next word, which gears it for general chat usage. Trinity Large Base is the base model without post-training. Then we have TrueBase, a model with any instruct data or post training so enterprises or researchers that want to customize it won't have to unroll any data, rules, or assumptions. Arcee AI will eventually offer a hosted version of its general-release model for 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 $0.045 / $0.15, and there is a rate-limited free tier available, too. Meanwhile, the company still sells post-training and customization options. Previously Arcee was doing model customization for enterprise clients like SK Telecom. The company was originally doing post-training: taking a Llama model, a Mistral model, a Qwen model that was open source, and post-training it to make it better for a company's intended use, including doing the reinforcement learning. But as their client list grew, the need for their own model was becoming 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 were barred from using.
It was a nerve-wracking decision. "I think there's less than 20 companies in the world that have ever pre-trained and released their own model" at the size and level that Arcee was gunning for, McQuade said. The company started small at first, trying its hand at a 4.5B model created in partnership with DatologyAI. The project's success then encouraged bigger endeavors. Previously Arcee was doing model customization for enterprise clients like SK Telecom. CTO Lucas Atkins and CEO Mark McQuade were quoted in the article. Arcee started with smaller models: Trinity Mini (26B) and Trinity Nano (6B).
