Trinity-Large-Thinking ships under Apache 2.0 for agent workloads
- Source
- x.com
- Date
Today we're releasing Trinity-Large-Thinking. Available now on the Arcee API, with open weights on Hugging Face under Apache 2.0. We built it for developers and enterprises that want models they can inspect, post-train, host, distill, and own.
Preview showed us where the demand was going. People were already running Trinity-Large-Preview in real agent workflows, with long-horizon tool use and production constraints. So over the last two months, we pushed our SFT and RL stack to meet that moment.
Our focus was clear. Build a model that stays coherent across turns, uses tools cleanly, follows instructions under constraint, and is efficient enough to serve at scale. That is the bet behind Trinity-Large-Thinking.
Context
Arcee released Trinity-Large-Thinking, available on its own API, on OpenRouter, and as on Hugging Face under the Apache 2.0 license, which permits commercial use, modification, and redistribution. Arcee frames the release around developers "that want models they can inspect, post-train, host, distill, and own," rather than only use through a hosted API.
The predecessor, Trinity-Large-Preview, was already seeing use in real agent workflows involving long sequences of tool calls under production constraints, according to Arcee, which it says prompted about two additional months of supervised and reinforcement learning aimed at that use case: staying coherent across many conversational turns, calling tools cleanly, following instructions under added constraints, and behaving predictably over long-running sessions.
Arcee's thread references a benchmark chart without stating the individual scores in the tweet text itself; readers who need the specific numbers behind the comparison should follow the linked launch blog. The model launched with a five-day free window on OpenRouter for OpenClaw and support at launch from coding and agent tools including Kilocode, Cline, and Opencode.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
- fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
- tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Checking sign-in…
Loading comments…