Arcee Trinity

Arcee Trinity is a family of open-weight, sparse Mixture-of-Experts (MoE) large language models developed by Arcee AI. Built for multi-turn conversations, tool orchestration, and structured data generation, Trinity models are distributed both as open weights for local or private cloud deployments via Hugging Face and as hosted, OpenAI-compatible API endpoints.

Model Architecture & Deployment Tiers

All Trinity variants utilize a sparse Mixture-of-Experts architecture where only a subset of total parameter experts activate per token, reducing inference latency and compute overhead. The model family shares a unified skill profile across three deployment footprints:

• Trinity Nano: 6B total parameters (1B active parameters per token) with a 128K-token context window, optimized for edge devices, embedded systems, and low-latency offline execution.
• Trinity Mini: 26B total parameters (3B active parameters per token) with a 128K-token context window, tuned for multi-turn agent orchestration and structured API outputs in cloud or on-premises backends.
• Trinity Large: 400B total parameters (13B active parameters per token) with a 256K-token context window, designed for long-horizon agentic loops, extensive document context utilization, and complex instruction following.

Capabilities, Tool Use, & Integrations

Trinity models natively support function calling, tool orchestration, and parallel tool execution, with built-in Model Context Protocol (MCP) support. The models generate schema-compliant JSON outputs and provide raw reasoning traces for auditing multi-step agent decisions. Trinity also serves as a streaming language model backbone for real-time voice applications, feeding low-latency text outputs into third-party text-to-speech (TTS) engines.

Training & Fine-Tuning

According to vendor documentation, Trinity models are trained on curated, domain-filtered datasets augmented with synthetic data specifically focused on tool calling, schema adherence, error recovery, and long-turn conversation coherence. Developers can inspect, fine-tune, or host Trinity weights directly using open-source LLM serving frameworks.