Local Models: Trust, Control, Optimization
Carter Abdallah · Vincent Weisser · Lucas Atkins · Chris Alexiuk
AI Engineer World's Fair 2026 · 43:21
← All organizationswww.arcee.ai ↗
Open-weight foundation models and model development tools
Arcee AI
Arcee AI builds Trinity open-weight foundation models and tools for developers and enterprises to customize language models. Its work has expanded from adapting existing models for specific domains to training its own foundation models. Trinity spans smaller Nano and Mini models through Large and Large-Thinking, with hosted API access and downloadable weights. Developers can inspect, post-train, distill and host the models themselves. The family uses OpenMDW-1.1, permitting modification and commercial deployment without a separate commercial license.
Founded in 2023, Arcee’s co-founders are CEO Mark McQuade, Brian Benedict and Jacob Solawetz. McQuade and Benedict previously worked at Hugging Face; Solawetz came from computer-vision company Roboflow. Its engineering contributions include MergeKit, an open-source library that combines model parameters to bring together capabilities from different checkpoints without additional training. Its Spectrum approach selectively trains useful layer modules while freezing others, reducing the amount of a model that needs updating during customization.
In 2026, the company reported that Trinity-Large-Preview served 3.37 trillion tokens through OpenRouter in its first two months. By June, it reported 30 employees, including 14 in research, and more than 200 models on Hugging Face with millions of downloads. That year, Arcee entered a multi-million-dollar partnership making Hugging Face the exclusive home for its models, datasets and agent traces.
2 talks
Newest first
2 speakers at AIE
Affiliations reflect their AIE appearances, not necessarily current employment.
Start here
- The Base Model is Dead
Start with Varun Singh’s comparison of model training strategies to understand how code, STEM, and reasoning data prepare models for downstream reinforcement learning.
Varun SinghAI Engineer World's Fair 2026
- Local Models: Trust, Control, Optimization
Turn to this panel for a discussion of collaboration around NVIDIA Nemotron and Arcee Trinity and the motivations for choosing local models over centralized APIs.
Carter Abdallah · Vincent Weisser · Lucas Atkins · Chris AlexiukAI Engineer World's Fair 2026
Messages from the stage
Post-training requirements move upstream
Varun Singh describes how synthetic data and supervised fine-tuning-style examples are entering pre-training, shifting base models away from primarily mirroring scraped web text.
Openness needs practical infrastructure
The panel featuring Lucas Atkins connects trust in model provenance with efficient inference, accessible hardware, and developer participation as conditions for advancing open models.
Affiliations reflect each recorded session, not necessarily current employment. The local models session is a joint discussion with panelists from NVIDIA, Prime Intellect, and Arcee AI.