Trinity Large Thinking

Trinity Large Thinking is a sparse Mixture-of-Experts reasoning model built around an AfmoeForCausalLM architecture, totaling roughly 400 billion parameters with about 13 billion active per token during inference. The model is explicitly designed for long-horizon planning, tool use, and multi-step agent workflows, emitting structured reasoning traces inside `<think ...` blocks so developers can inspect how conclusions were reached. Its design as a frontier-class open-weight model reflects Arcee's intent to offer a Western alternative to the Chinese open models that have dominated the reasoning space, enabling both on-premises deployments and API access for teams who want full control over their inference environment.

The model was post-trained using extended chain-of-thought reasoning and agentic reinforcement learning to cultivate the kind of multi-turn deliberation that real-world agent tasks demand. Multiple GPU configurations are documented in the wild—vLLM recipes reference tensor-parallel setups across 8 GPUs for production latency, and community discussions suggest it can compress into two DGX Spark nodes. Quantized GGUF variants broaden accessibility for users running on constrained hardware, while the Apache 2.0 license removes deployment barriers for enterprises building autonomous workflows. The result is a model tuned for agents that must plan several steps ahead, call external tools reliably, and maintain coherent context across extended conversations.