According to 1M AI News, American AI model company Arcee has released Trinity-Large-Thinking, an open-source reasoning model designed for long-duration Agent tasks. The model utilizes a sparse mixture of experts (MoE) architecture, with a total of 400 billion parameters and only 13 billion active parameters, available for download under the Apache 2.0 license on Hugging Face. Unlike its predecessor Trinity-Large-Preview (pure instruction fine-tuning), Trinity-Large-Thinking performs reasoning before answering, showing improvements in multi-round tool invocation, long context coherence, and instruction-following capabilities, with the core design goal of maintaining stable output during long-duration Agent cycles. It scored 91.9 on the Agent capability benchmark PinchBench, ranking second, just behind Opus 4.6's score of 93.3; on the Agent task benchmark Tau2-Airline, it scored 88.0, the highest among all comparison models. However, its performance on general reasoning benchmarks was average: it scored 76.3 on GPQA-D, lower than Kimi-K2.5 (86.9) and Opus 4.6 (89.2); and 83.4 on MMLU-Pro, also ranking last. According to Arcee, the model is "the strongest open-source model outside of China on many dimensions." The Arcee API is priced at $0.90 per million tokens, which Arcee claims is about 96% cheaper than Opus 4.6. The model is also available on the AI model routing platform OpenRouter, where it can be used for free in OpenClaw for the first five days. Since its release at the end of January, the previous Preview has served over 33.7 trillion tokens on OpenRouter, making it the most used open-source model in the U.S. and the fourth globally, with Preview continuing to be offered for free on OpenRouter.