Trinity Large

Trained entirely in the United States by startup Arcee AI, the Trinity family argues that sovereign open models can compete at frontier scale. Trinity-Large packs about 400B total parameters into an unusually sparse mixture-of-experts layout, activating 4 of 256 experts for roughly 13B parameters per token, a 1.56 percent routing fraction that yields 2 to 3x the throughput of comparably sized dense models, with a 256K context window. The reasoning variant Trinity-Large-Thinking, released April 1, 2026 under Apache 2.0, adds extended chain-of-thought post-training and agentic reinforcement learning, scoring 91.9 on the PinchBench agentic benchmark, just behind the strongest proprietary systems. Smaller siblings Trinity-Nano (6B total, 1B active) and Trinity-Mini (26B total, 3B active) cover edge and mid-range deployment. Weights live on Hugging Face and run under vLLM, SGLang, llama.cpp, LM Studio, and Transformers, with managed APIs on Arcee's platform and OpenRouter; Apache terms permit unrestricted enterprise use.