Arcee AI drops Trinity-Large-Thinking open reasoning model

Trinity-Large-Thinking is a frontier open reasoning model from Arcee AI, built specifically for complex, long-horizon agents and multi-turn tool calling, and released fully under the Apache 2.0 license.

If you have been running autonomous agents 24/7, you know the drill: models lose context coherence after a few turns, tool calling breaks, and long agent loops get sloppy. This release takes aim squarely at those failure modes.

What it actually does

Unlike its predecessor Trinity-Large-Preview, which was an instruct checkpoint, Trinity-Large-Thinking adds a deliberate reasoning step before it generates any final response.

Basically, it thinks through the task before reaching for external tools. That single design choice yields much tighter instruction following under constraints, dependable multi-turn tool execution, and far more consistent behavior across sprawling agent loops.

The benchmark numbers back it up. On PinchBench, an agent capability benchmark from Kilo, Trinity-Large-Thinking ranks number two in the world, sitting right behind Opus-4.6.

Then there is the economics. On the Arcee AI API, output tokens cost just $0.90 per million. That makes it roughly 96% cheaper than Opus-4.6, fundamentally changing what it costs to let agents run unsupervised for hours.

Hardware and specs under the hood

Arcee AI stood up this pipeline by leaning heavily into hardware efficiency across pretraining and RL.

The base model was pretrained on 2,048 NVIDIA B300 GPUs, followed by post-training across 1,152 NVIDIA H100s to align the reasoning pipeline. In production, their serving stack runs on NVIDIA Dynamo, Blackwell Ultra GPUs, and vLLM.

The model builds directly on real-world mileage. The earlier Preview version served more than 3.37 trillion tokens on OpenRouter in its first two months, becoming the top-used open model in the United States on the OpenClaw collection.

Openness and where to run it

This is a genuinely permissive release. The model weights are live on Hugging Face under the Apache 2.0 license, meaning enterprises can inspect, self-host, fine-tune, or distill them without restrictive commercial terms.

If you want to plug it into your stack immediately, it is available through Arcee AI's API platform as well as OpenRouter. The original Preview checkpoint will also remain free on OpenRouter, though shifted to a lighter compute pool.

Arcee also plans to take the lessons from this model to distill smaller Trinity-2-Nano and Mini variants down the road.

Why it matters

If you are an engineer or startup building multi-step agentic workflows, Trinity-Large-Thinking should be in your testing pipeline today. You no longer have to choose between burning through API budgets on top-tier proprietary models and settling for open weights that fall apart during tool calling.

We think Arcee played this smart by skipping the general coding hype and obsessing over the messy reality of agent loops. Providing near-frontier performance under Apache 2.0 at under a dollar per million tokens sets a brutally competitive new floor for both proprietary vendors and open-weight alternatives.