Arcee AI releases Trinity Large Thinking, an open reasoning model for AI agents

On April 1, 2026, Arcee AI released Trinity Large Thinking, an open reasoning model designed not for short chat replies but for long agentic workflows that involve multiple steps, tool calls, and context that carries over between turns. That matters for the market: in a segment where closed models from the big labs set the pace, there is now an open option under the Apache 2.0 license. Developers can call it through an API, but they can also run it, fine-tune it, and build it into their own systems without any licensing gray areas.

Trinity Large Thinking is the reasoning version of the Trinity Large family, which Arcee is developing as an open alternative to large proprietary models. The company published the weights on Hugging Face and launched the model in its own API at the same time. The bet here is not on a "universal chatbot for everything" but on tasks where an agent has to hold a plan, remember earlier steps, use tools carefully, and not fall apart after a few iterations. These are exactly the scenarios that are becoming the baseline for AI development: coding agents, operators for internal systems, enterprise assistants, and pipelines that make multiple calls to external services.

Architecturally, the model is a sparse MoE: Trinity Large Thinking has around 398-400 billion parameters, but only about 13 billion are active per token. It contains 256 experts, of which only four run at once. The design aims to keep a high quality ceiling without making inference completely prohibitive. Arcee also says the model supports up to 512 thousand tokens of context after window extension, which matters especially for long agent loops, large repositories, extensive documentation, and complex multi-turn tasks. One more detail: the model generates an explicit "thinking layer" before it answers, and developers are advised to preserve this reasoning context between turns, since otherwise multi-step performance can drop noticeably.

The most interesting part of the release is not just the license but the stated focus on practical agentic work. According to Arcee, Trinity Large Thinking ranks second on PinchBench, behind only Claude Opus 4.6, and is noticeably stronger than the earlier Trinity Large Preview specifically at multi-step tool use, instruction following, and staying coherent over long runs. In the model card, the company also reports strong results on a number of agentic benchmarks, including τ²-Bench and LiveCodeBench. Arcee is also leaning on economics: at announcement time, the company put the inference cost at roughly $0.90 per million output tokens, pitching the model as a substantially cheaper alternative to closed reasoning systems for production agents.

The context around the release matters too. Trinity Large Preview, introduced in late January 2026, served 3.37 trillion tokens through OpenRouter in its first two months, according to the company. Arcee claims the preview version became the most-used open model in the US in the OpenClaw collection and fourth globally. For a small team, that is a way to show that demand for open models in real agentic workloads already exists, not at the demo level but as sustained production load. Technically, the project was no small feat either: Trinity Large was pretrained on 17 trillion tokens using 2048 NVIDIA B300 GPUs, and the company previously estimated the entire path to the Large family at about $20 million.

The main takeaway from the release: the open-source AI market is shifting from a race over "who writes the best text" to a race over "which model handles long-running work most reliably." Trinity Large Thinking matters not because it instantly beat the closed leaders on every metric, but because it gives developers and companies another genuinely open option for building agentic systems without API-only restrictions. The question now is not whether an open reasoning model of this class can be released, but how consistently it will perform in production, where what counts is not polished demos but hours-long loops, the cost of mistakes, and predictable behavior.