Arcee AI: Trinity Large Thinking

Arcee AI: Trinity Large Thinking is a text model from Arcee AI. It is an open source reasoning model, which means the weights and the reasoning behaviour behind it are published rather than kept proprietary. Arcee AI positions it as a large model built specifically for thinking tasks, and the provider reports strong performance in PinchBench, agentic workloads, and reasoning tasks. If you are choosing a model for work that involves multi step deliberation rather than quick single pass generation, this is the family it belongs to.

The model accepts text and returns text. Its context window is 262144 tokens, which is large enough to hold long documents, extended conversation histories, or sizeable codebases in a single request. The maximum output is 80000 tokens, so responses can be long when the task calls for it. Because it is a reasoning model, it can expose its reasoning as part of the response, and you control that behaviour through the reasoning and include_reasoning parameters rather than getting it unconditionally.

Those capabilities point to a fairly specific set of uses. Long context plus reasoning suits agentic workflows where the model has to plan, call tools, inspect results, and continue over many steps. It fits analysis of long documents, such as contracts, research papers, or logs, where the answer depends on details spread across a lot of text. It also fits tasks that benefit from visible deliberation, since you can request the reasoning trace and inspect how a conclusion was reached. Tool use is supported, so function calling and structured agent loops are viable.

You call it through api14's OpenAI-compatible API using the model id "arcee-ai/trinity-large-thinking". Because the interface follows the OpenAI schema, existing clients and SDKs work with a change of base URL and model id. The supported request parameters are include_reasoning, max_tokens, reasoning, temperature, tool_choice, tools, top_k, and top_p. In practice that covers output length control, sampling behaviour through temperature, top_p, and top_k, tool definitions and selection, and control over whether reasoning is returned.

One thing to keep in mind is that reasoning models trade latency and token spend for answer quality, so for simple classification, short summarisation, or high volume low complexity generation, a smaller non reasoning model may be the more economical choice. The model is text only, so image, audio, or other modalities are out of scope here. Within those bounds, the combination of a 262144 token context, an 80000 token output ceiling, and tool support makes it a reasonable candidate for agentic and long document work.