Arcee, a tiny 26-person startup based in the U.S., just released Trinity Large Thinking, a new reasoning model that CEO Mark McQuade is calling the most capable open-weight model ever released by a non-Chinese company. And honestly? There’s something genuinely exciting about rooting for the underdog here.

Here’s why this matters: Arcee built a massive 400-billion-parameter open source language model on just a $20 million budget. That’s the kind of efficiency that makes you wonder what we’ve all been doing with our venture capital. The company’s mission is straightforward, give Western companies a legitimate alternative to Chinese AI models, which many see as risky given data privacy concerns and geopolitical tensions.

Unlike the closed-source models from Anthropic or OpenAI, Arcee’s Trinity models are actually yours to download, train on your own terms, and run on your own servers. You can also use their cloud-hosted version through an API if that’s more your speed. This isn’t just a technical difference, it’s philosophical. You’re not beholden to the whims of corporate overlords suddenly changing their pricing or usage policies.

Which brings us to a perfect example of why open source matters. Anthropic recently told users of the OpenClaw tool that their Claude subscriptions no longer cover that usage, they’ll need to pay extra. It’s the kind of move that frustrates developers who’ve built their workflows around a particular model. Arcee, meanwhile, has become one of the top models used with OpenClaw, according to data from OpenRouter. McQuade is rightfully proud of that.

Now, let’s be real about the benchmark results. Trinity Large Thinking is competitive with other top open source models, but it’s not necessarily outperforming the closed-source giants from OpenAI or Anthropic. It’s also not the head-to-head threat to Meta’s Llama 4 that some might hope for. But here’s what sets Arcee apart: all their Trinity models are released under Apache 2.0, which is considered the gold standard for open source licenses. Meta’s Llama, despite its marketing, actually has some licensing weirdness that makes it not truly open source in the way Arcee’s models are.

The broader picture here is that there’s a whole ecosystem of U.S. startups building open source AI models, and yeah, we’re rooting for all of them. But Arcee’s approach, combining genuine accessibility with legitimate technical capability, represents something important. It’s a reminder that you don’t need infinite resources and corporate backing to make meaningful contributions to AI. Sometimes you just need smart people with a clear mission and the determination to pull it off on a shoestring budget.