Arcee, a San Francisco-based AI startup, has introduced Trinity-Large-Thinking, a 399-billion parameter AI model released under the Apache 2.0 license for full commercial use and customization. This model uses an advanced sparse Mixture-of-Experts architecture, activating just 1.56% of its parameters per token, achieving high reasoning capabilities and operational efficiency. Backed by a $24 million Series A round and a $20 million investment in training, Arcee embodies “engineering through constraint” with a compact 30-person team. Trinity-Large-Thinking shifts from “instruct” to “reasoning” models, making it adept at complex agentic and multi-step tasks with faster, more coherent outcomes. Its data training leverages 20 trillion tokens combining curated web and synthetic rewritten content to promote reasoning rather than rote memorization, ensuring regulatory compliance critical for enterprise use. Arcee aims to fill growing demand for U.S.-based open AI models, as global competitors shift toward proprietary platforms. Benchmark scores show Trinity competing strongly against other top open and closed-source models, offering a cost-effective and sovereign option for enterprises building autonomous agents. Apache 2.0 licensing empowers enterprises with true ownership, crucial for regulated industries; Arcee also offers a raw pre-trained checkpoint for transparency and customization.
Back
