Arcee AI has stunned the AI world by releasing Trinity, a 400-billion-parameter open-source large language model (LLM) built from scratch, claiming performance comparable to Meta's Llama 4 Mave on key benchmarks like coding, math and reasoning. What makes the launch remarkable is that this achievement comes from a startup of about 30 people, trained in just six months using efficient hardware and a $20 million compute spend.

Unlike many high-end models tied to proprietary or restricted licenses, Trinity is published under a permanently open Apache 2.0 license, ensuring free access to developers, researchers and enterprises. Arcee emphasizes this licensing choice as a strategic differentiator, positioning Trinity as a genuinely open alternative to models with usage caveats. The startup plans three Trinity variants — from a base model for customization to a lightly post-trained instruct model geared for general chat use — with hosted API services coming soon.

The Trinity family builds on Arcee’s earlier scaled-down releases like Trinity Mini (26B) and Trinity Nano (6B), showing a clear roadmap from efficient miniature models to frontier-class AI. With competitive benchmarks and free access, Arcee AI aims not only to challenge big tech incumbents but also to invigorate the open-source AI ecosystem by lowering barriers to cutting-edge foundation models.