Trinity Mini

Trinity Mini is Arcee's compact sparse Mixture-of-Experts model with 26 billion total parameters and 3.5 billion active parameters per token, optimized for resource-efficient inference in production agent systems. Released as open weights under the Apache 2.0 license, it preserves strong capability while sharply reducing computational overhead compared with the larger 400B Trinity Large family. The model features a 128k context window with grouped query attention and supports function calling and structured output, enabling multi-turn agent workflows and tool orchestration while maintaining coherence over long inputs. Trained on 10 trillion tokens, Trinity Mini delivers performance close to larger instruction-following models while prioritizing speed and reliability in agentic loops, with recommended inference settings tuned for low-latency multi-step work. Unlike Trinity Large Preview, which leans toward creative tasks, or Trinity Large Thinking, which optimizes for explicit reasoning, Trinity Mini targets developers who need practical, lightweight agentic capability. Its smaller parameter count and MoE efficiency make it well suited to applications requiring rapid token generation and memory-efficient multi-turn conversations.