Arcee, a compact 26-person startup based in the U.S., has unveiled its latest reasoning model. This team managed to develop a substantial 400-billion-parameter open source large language model with a modest budget of just $20 million. Their underlying mission is one that many find compelling: to provide U.S. and Western companies with a capable alternative that eliminates any need to rely on Chinese-based models. Although Chinese models are highly advanced, they are often viewed as carrying inherent risks, potentially placing both influence and data under the control of a government whose values differ from those widely held in the West. With Arcee’s offering, businesses have the flexibility to download the model, customize it through training to meet their specific requirements, and deploy it entirely on their own infrastructure. Alternatively, companies can opt for Arcee’s cloud-hosted version, accessible via an API.

While Arcee’s models may not yet surpass the performance of closed-source alternatives from industry leaders like Anthropic or OpenAI, they offer a distinct advantage: freedom from the unpredictable policy changes of those larger entities. A recent example involves Claude, a model renowned for its coding proficiency, which had become a favorite among users of the open source AI agent platform OpenClaw. Last week, however, Anthropic unexpectedly changed course, informing users that their existing subscriptions would no longer cover OpenClaw usage, requiring an additional payment. This move followed OpenClaw creator Peter Steinberger’s announcement in February that he was joining OpenAI, Anthropic’s primary competitor. In a contrasting display of stability, Arcee’s CEO Mark McQuade highlights data from OpenRouter indicating that their model has risen to become one of the most frequently used with OpenClaw. This naturally leads to the question: just how capable is Trinity Large Thinking?

As noted in earlier coverage, this model is not positioned as a direct competitor to the leading U.S.-built open model, Meta’s Llama 4. However, it also avoids the complicated and somewhat restrictive licensing terms associated with Meta’s offering. All of Arcee’s Trinity models are released under the Apache 2.0 license, widely regarded as the gold standard for open source permissions. It is important to acknowledge that Arcee is among a vibrant ecosystem of countless other U.S. startups developing open source models. As an admirer of startup innovation, I find their collective ambition equally worthy of support.