China's strong presence in the open weight language models segment has become undeniable. While it is true there are commercially available options made by US companies, their status seems to be on shaky ground: OpenAI released two open-weights reasoning models last year, but for obvious reasons, would much rather have customers buy into its closed-source offerings. Meta has become the go-to, but even there the future is uncertain, given the company has hinted that, going forward, it may withhold from publishing its most advanced technology under open source licensing.

It is perhaps for these reasons that Arcee AI, a 30-person, US-based tiny startup, is betting big on creating truly state-of-the-art open source foundation models fully developed in the US. The company just released Trinity Large, a 400-billion-parameter model under the permanently open Apache license—one of the largest open source foundation models ever trained by a US company. According to benchmark tests, it performs comparably to Meta's Llama 4 Maverick and China's high-performing GLM-4.5 from Tsinghua University.

The achievement is remarkable given that Arcee AI never meant to be a full-fledged AI lab. Arcee's business relied exclusively on post-training models to customize them for the needs of enterprise customers like SK Telecom. As the company's client list grew, so did the need for a fully local model; one that would allow Arcee to stop relying on other companies' models, and that would allay customers concerns about using models from Chinese labs. "The US needs a permanently open, Apache-licensed, frontier-grade alternative," said founder and CEO Mark McQuade, formerly an early Hugging Face employee.

Even more impressive is the fact that Arcee AI trained Trinity in just six months for $20 million using 2,048 Nvidia Blackwell B300 GPUs—a fraction of what larger labs spend. Benchmarks show Trinity matching or slightly beating Llama on coding, math, common sense, and reasoning tasks. Given the constraints, the Trinity models are currently text-only, but the company says that a model with vision capabilities is in the works, and that its roadmap contemplates breaking into the speech-to-text segment.

Trinity is available for free download, with a competitively priced hosted version expected within six weeks.