Published: April 3, 2026
Arcee AI, a San Francisco-based open-intelligence lab, has released Trinity-Large-Thinking, a 399-billion parameter reasoning model licensed under Apache 2.0. The model is a sparse Mixture-of-Experts architecture with roughly 13 billion active parameters per token, making it one of the most capable open-source models available for enterprise customization.
Architecture and training
Trinity-Large-Thinking is built on the Trinity-Large-Base, a 398-billion parameter MoE model using 256 experts with 4 active per token. The model has an unusually high sparsity ratio of 1.56% and uses 6 dense layers (increased from an initial 3) to maintain routing stability across experts.
Training consumed $20 million in resources, nearly half of Arcee’s total funding, and ran for 33 days on a cluster of 2,048 NVIDIA B300 Blackwell GPUs. The base model was trained on 17 trillion tokens, with over 8 trillion being synthetic data. DatologyAI handled data curation, covering programming, STEM, reasoning, and 14 non-English languages.
Trinity-Large-Thinking adds extended chain-of-thought reasoning and agentic reinforcement learning on top of the base model. It generates explicit reasoning traces wrapped in think blocks before producing final responses, similar to the approach used by other reasoning-focused models.
Performance
The Trinity-Large Preview variant scores 87.2 on MMLU and 24.0 on AIME 2025. Arcee claims the model matches or exceeds peers in math, coding, and scientific reasoning tasks, while running roughly 2 to 3 times faster on equivalent hardware due to its sparse activation pattern.
The model supports a 512,000-token native context window, which positions it well for long-document analysis and extended agentic workflows that require maintaining context over many turns of interaction.
Open-source strategy
Trinity-Large-Thinking is released under the Apache 2.0 license, allowing unrestricted commercial use and full customization. Arcee describes the model as a rare example of a powerful, U.S.-made AI model that enterprises can download and modify without licensing restrictions. The full model weights are available on Hugging Face in three variants: Preview (instruction-tuned), Base (full 17T token checkpoint), and TrueBase (early 10T token checkpoint with no instruct data).
Availability
The model is live on Hugging Face and available through OpenRouter. It is also deployed on DigitalOcean’s Agentic Inference Cloud for managed hosting.
Read the full technical details on Arcee AI’s blog.
