Arcee AI has launched the Trinity-Large-Thinking model, an open reasoning AI designed to excel in agent tasks, competing directly with Claude Opus. The project, consuming nearly $20 million or half of the company’s venture capital, features 400 billion parameters and employs a mixture-of-experts architecture, activating about 13 billion parameters per token for efficient performance. Although it shows competitive results in agent benchmarks like Tau2 and PinchBench, Trinity-Large-Thinking underperforms in general reasoning assessments like GPQA-D and MMLU-Pro.
Built on 2,048 Nvidia B300 GPUs over 33 days, the model demonstrates capability in handling extensive context with a 512K token window. The innovative training strategy, including synthetic data generation and a custom loading balance method, optimized performance without spikes in training loss. Early adoption has been strong, with the model actively utilized on platforms such as OpenRouter. Future iterations will include further fine-tuning to enhance its capabilities.
Source link
