Exploring “Trinity-Large-Thinking: Scaling an Open Source Frontier Agent” with Lucas and Varun

today we have the pleasure of chatting with two core contributors of the open-weight frontier model Trinity from Arcee: Lucas Atkins and Varun Singh Trinity-Large-Thinking has just released a few weeks ago and it is an open-weight (apache 2.0) reasoning model built on a 400B/13B-active base, purpose-built for long-horizon agents and landing at #2 on PinchBench behind Claude Opus 4.6 at roughly 96% lower cost. we’re gonna go through: a high-level walkthrough of Trinity Large: high sparsity MoE, SMEBU, Muon optimizer, Random Sequential Document Buffer, and the 6 last-minute changes that shaped the final run how Arcee, Datalogy, and Prime Intellect coordinated a multi-lab training run across three organizations lessons going from AFM-4.5B (dense, solo-infra on AWS) to an extremely sparse MoE at scale. what's next for Trinity: mid-training phases, the potential path to multimodal, and the short-term roadmap come hang out and share your questions with the authors (and thanks for everyone that submitted theirs)! 🌹