Trinity Large Preview

Trinity Large Preview is Arcee's 400B-parameter sparse Mixture-of-Experts model released in January 2026, featuring 13 billion active parameters per token and a 256-expert architecture with 4 experts active per token. It serves a 128k context window and was trained on 17 trillion tokens curated across programming, STEM, reasoning, and multilingual data. This preview variant is positioned for creative writing, storytelling, role-play, chat, and real-time voice scenarios, alongside multi-step agent workflows. Arcee trained it to navigate agent harnesses and to handle complex toolchains and long, constraint-filled prompts, making it well suited to conversational and tool-augmented applications. Trinity Large Preview is released as open weights under the Apache 2.0 license. It is the instruct-tuned sibling of Trinity Large Thinking, leaning toward production-grade conversational and creative output rather than explicit step-by-step reasoning traces.