Arcee AI Releases Trinity-Large-Thinking Open Reasoning Model

U.S.-based startup Arcee AI has released its new open reasoning model, Trinity-Large-Thinking, announced in a company blog post. The model is available through Arcee’s API and on Hugging Face under the Apache 2.0 license.

Trinity-Large-Thinking is designed for complex, multi-turn agent tasks and long-horizon tool calling. It builds on the earlier Trinity-Large-Preview model, introducing improved context coherence, instruction following, and stability across extended agent runs. The company reports that the model ranks second on the PinchBench benchmark for agent-related tasks, behind Opus-4.6, while operating at a lower cost per token.

Arcee trained the model using 2048 NVIDIA B300s for pretraining and 1152 H100s for post-training. The company says its optimized pipeline now powers inference using NVIDIA Dynamo and Blackwell Ultra GPUs. According to CEO Mark McQuade, Trinity-Large-Thinking represents a significant step toward producing high-quality, open-weight models that developers and enterprises can train, host, and modify independently.

The release follows the success of Trinity-Large-Preview, which became one of the most used open models on OpenRouter’s OpenClaw collection. Arcee plans to apply lessons from this model to future versions of its smaller Trinity models, including Mini and Nano.