MODELS

Arcee AI: Trinity Large Thinking

by Arcee AI

modellead-sourceopenrouter-modelsreasoning-modelopen-weightstool-callingagentssparse-moearcee-aisource:arcee.aistructured-outputshugging-faceopenrouter

Visit Website Documentation

Overview

Open reasoning model from Arcee AI for long-horizon agents, multi-turn tool calling, and reasoning tasks.

Details

Trinity-Large-Thinking is Arcee AI’s reasoning-optimized Trinity-Large variant. Arcee documentation and the Hugging Face model card describe it as a 398B sparse MoE model with approximately 13B active parameters per token, built on Trinity-Large-Base. Arcee positions it for long-horizon agents, multi-turn tool calling, and agentic/tool-calling use cases, with agentic RL post-training. The Trinity product page lists Trinity Large Thinking as served on the Arcee API with 512K context and capabilities for tool use, structured outputs, and multi-turn conversations. Arcee’s announcement says it is available on the Arcee API with Hugging Face weights under Apache 2.0.

When to Use

Use when evaluating an open reasoning model for long-horizon agent workflows. Use for multi-turn tool-calling experiments where agentic behavior and reasoning are primary requirements. Use when Apache 2.0-licensed model weights are important to your evaluation or deployment process. Compare it with other reasoning-optimized models for agentic workloads and structured-output workflows.

Getting Started

Read the Trinity-Large-Thinking documentation page for architecture, post-training, tool-calling focus, and licensing details.
Review the Trinity product page to understand Arcee API availability, context length, tool use, structured outputs, and multi-turn conversation support.
Use the Arcee chat-completions API reference for model selection, streaming, tools, response_format, and reasoning_effort parameters.
Check Arcee’s pricing page for current text-token pricing before running workloads.
Evaluate the Hugging Face model card if you want to inspect or use the released weights.

Key Features

•398B sparse MoE reasoning-optimized Trinity-Large variant, according to Arcee documentation and the Hugging Face model card.
•Approximately 13B active parameters per token.
•Agentic RL post-training with a tool-calling focus.
•Positioned for long-horizon agents and multi-turn tool calling.
•Listed by Arcee with 512K context on the Trinity product page.
•Supports tool use, structured outputs, and multi-turn conversations on the Arcee API, according to the Trinity product page.
•Hugging Face weights are released under Apache 2.0, according to Arcee’s announcement and documentation excerpts.

Capabilities

•reasoning
•long-horizon agents
•multi-turn tool calling
•tool use
•structured outputs
•multi-turn conversations
•agentic workloads
•sparse MoE language modeling

Last updated Jun 2, 2026