Arcee, a small U.S. startup with just 26 employees, has developed a massive 400-billion-parameter open-source LLM on a modest $20 million budget. This model, called Trinity Large Thinking, is touted by CEO Mark McQuade as the most capable open-weight model ever released by a non-Chinese company.
Arcee aims to provide Western companies an alternative to Chinese models, which, despite their capabilities, are often perceived as risky due to concerns about data security and differing governmental ideals.
Arcee’s model can be downloaded, trained for specific needs, and used on company premises. There’s also a cloud-hosted version available via API. While it doesn’t outperform leading models from Anthropic or OpenAI, it offers independence from their influence.
For example, users of the open-source AI agent tool OpenClaw, which previously favored Anthropic’s Claude for its coding abilities, have faced new costs as Anthropic now requires additional payment for OpenClaw support. Meanwhile, McQuade highlights data from OpenRouter indicating Arcee’s model has become a top choice for OpenClaw.
Trinity Large Thinking is competitive with other leading open-source models, according to benchmark results. While not directly challenging Meta’s Llama 4, it avoids Meta’s licensing issues by being released under the Apache 2.0 license, a gold standard for open-source licenses.
Numerous other U.S. startups are also offering open-source models, and support for the innovation of these companies continues to grow.