It’s hard not to root for Arcee, a small, open-source AI model maker.

Arcee, a 26-person US startup that built a large-scale 400B parameter open-source LLM on a shoestring budget of $20 million, has released a new inference model. Arcee calls this model the Trinity Large Thinking, and it’s the most capable open-weight model “released by a company outside of China,” CEO Mark McQuade claims to TechCrunch.

As you can tell from that comment, Arcee has a goal that I can’t help but root for. Arcee wants to provide U.S. and Western companies with a model where there is no reason to use China-based products.

The Chinese model is very capable, but it is perceived as dangerous, putting power and data in the hands of a government that does not share all of the ideals of the Western world.

Arcee allows companies to download models, train them as needed, and use them on-premises. Businesses can also use a cloud-hosted version of Arcee, accessible via API.

Arcee’s models don’t outperform closed-source models from large labs like Anthropic or OpenAI, but they’re also not held hostage to the whims of giant corporations.

For example, Claude, with his strong coding skills, was a popular choice among users of OpenClaw, an open source AI agent tool. But Anthropic withdrew this last week, telling users that their Anthropic subscription would no longer guarantee use of OpenClaw. You will have to pay extra for this. (Last February, OpenClaw founder Peter Steinberger announced he was joining Anthropic’s biggest rival, OpenAI.)

By contrast, McQuade proudly points to data from OpenRouter, which he says has made it one of the best models used with OpenClaw.

So how good is Trinity Large Thinking? Benchmark results shared with TechCrunch show comparisons to other leading open source models.

As we previously reported, this doesn’t pose a direct threat to Meta’s Llama 4, a giant among US-made public models. But it also doesn’t have the weird, not-really-open-source licensing issues of the Meta model. All Trinity models from Arcee are released under Apache 2.0, the standard for OS licensing.

To be clear, there are plenty of other American startups offering open source models, and as a fan of startup creativity, I’m rooting for them too.