Trinity-Large-Thinking by Arcee AI is a 398B-parameter sparse Mixture-of-Experts (MoE) model with 13B active parameters per token, specifically optimized for reasoning and agentic workflows. It is post-trained with extended chain-of-thought reasoning and agentic RL, achieving strong performance on agentic benchmarks like "τ²-Bench" (94.7%), "PinchBench" (91.9%), and "LiveCodeBench" (98.2%). This model generates explicit reasoning traces within ... blocks, which are crucial for multi-turn conversations and agentic loops, and supports a 512k context length after extension.
Loading preview...
Trinity-Large-Thinking: Reasoning-Optimized Agentic MoE
Trinity-Large-Thinking is a 398-billion parameter sparse Mixture-of-Experts (MoE) model from Arcee AI, featuring approximately 13 billion active parameters per token. It is a specialized variant of the Trinity-Large family, post-trained with extended chain-of-thought reasoning and agentic Reinforcement Learning (RL) to excel in complex, multi-step tasks.
Key Capabilities
- Agentic-first design: Purpose-built for tool calling, multi-step planning, and sophisticated agent workflows.
- Native Reasoning Traces: Generates explicit reasoning within
<think>...</think>blocks, which must be preserved in context for effective multi-turn and agentic operations. - High Agentic Performance: Achieves 94.7% on τ²-Bench, 91.9% on PinchBench, and 98.2% on LiveCodeBench, demonstrating state-of-the-art results in agentic benchmarks.
- Broad Compatibility: Works seamlessly with major agent frameworks like OpenClaw and Hermes Agent.
- Extended Context Window: Supports a 512k context length, accommodating long reasoning chains across many agentic steps.
Good for
- Developing AI agents requiring robust tool-calling and multi-step planning capabilities.
- Applications that benefit from explicit, interpretable reasoning traces.
- Complex, multi-turn conversational systems where maintaining chain-of-thought is critical.
- Integration into existing agent frameworks like OpenClaw and Hermes Agent for enhanced performance.