Arcee AI unveiled Trinity Large, a new 400-billion-parameter sparse Mixture-of-Experts (MoE) foundation model designed to deliver frontier-class performance, efficiency, and versatility for developers, researchers, and enterprises. Trinity Large expands the company’s open-weight Trinity model family and makes advanced generative AI more accessible and practical for real-world use cases.
Trinity Large features a high-sparsity architecture with 256 experts and 4 active experts per token, providing an efficient balance of model capacity and inference performance. The model family is being released in three distinct checkpoints to support diverse developer needs and research workflows: Trinity-Large-Preview, Trinity-Large-Base, and Trinity-Large-TrueBase.
Trinity-Large-Preview is lightly post-trained and optimized for chat, creative writing, interactive storytelling, and agent use cases, offering strong performance for real-time conversational experiences. Trinity-Large-Base represents the fully pretrained checkpoint after processing 17 trillion tokens, demonstrating leading benchmark performance across a wide range of reasoning, math, coding, and knowledge tasks. Trinity-Large-TrueBase is an early checkpoint from the same run at 10 trillion tokens of pretraining without instruction data or learning-rate annealing, providing researchers with a true base model for in-depth analysis and fine-tuning.
Also Read: Cymulate Partners with Wiz Integration Network (WIN)
Arcee AI trained Trinity Large on 2048 NVIDIA B300 GPUs with an efficient, high-sparsity strategy that enabled training to complete in 33 days, balancing speed and computational cost while optimizing throughput and inference efficiency.
The Trinity Large architecture supports extended contexts with natively scalable context windows, and the introduction of high sparsity and optimized attention mechanisms enables performance improvements of roughly 2-3× on throughput compared to peers in the same weight class.
The company is releasing Trinity Large checkpoints under open-weight, permissive licensing, making them available through OpenRouter and Hugging Face. Preview models are free during the preview period, with integrations available at launch with partners including Kilo Code, Cline, and OpenCode to support diverse application scenarios without infrastructure overhead.
Arcee AI emphasized its commitment to open-source AI innovation and interoperability, enabling developers to run and experiment with Trinity Large across platforms and build robust multi-turn agents, structured outputs, and tool-driven workflows.
SOURCE: Arcee AI