Trinity Large Thinking

Trinity Large Thinking is Arcee's post-trained reasoning variant of the Trinity Large family, released in April 2026 and optimized for agentic workflows and autonomous agent systems. This roughly 398B-parameter sparse Mixture-of-Experts model activates around 13 billion parameters per token and supports a 256k context window, enabling complex multi-turn agent loops and reasoning chains across extended interactions. The model generates explicit reasoning traces wrapped in think blocks before producing its response, a design pattern that helps it maintain coherence in multi-step planning and long-horizon tasks. Built on the Trinity Large base with additional reinforcement learning, it scores 91.9 on PinchBench and 96.3 on AIME25, with strong tool calling and stable instruction following across long-running agent loops. Released as open weights under the Apache 2.0 license, Trinity Large Thinking differs from the Preview variant by prioritizing reasoning capability over general-purpose and creative output, making it suited to developer agents, autonomous task execution, and systems that need reliable agentic reasoning at scale.