Building Frontier Open Reasoning Models | Lucas and Varun (Arcee AI)

Arcee AI went from post training company to building their own frontier reasoning model from scratch + and open-sourcing all of it. In this episode, Lucas Atkins (Co-Founder & CTO) and Varun Singh (Pre-Training Lead) take us through the full story — the decision to start pre-training, how they built and trained Trinity on a fraction of the budget that frontier labs spend, the data and engineering challenges along the way, and their views on where open-source AI is headed. A really honest conversation about what it actually takes to build at the frontier.

TIMESTAMPS

00:00:00 - Intro

00:00:59 - Varun’s Transition from SWE to Pre-Training Lead

00:04:20 - The Trinity Manifesto & OpenClaw Ecosystem

00:12:15 - Arcee’s Post-Training to Pre-Training Pivot

00:23:45 - Varun’s First Pre-Training Run (“You Can Just Do Things”)

00:27:33 - Saturation in Pre-Training? Mid-Training Explained

00:37:00 - Tweaking Training Architecture, Adam vs Muon & Evals

01:09:07 - Inference Engineering, Quick Fire & Post-Training Recipes

01:18:02 - Alpha in RL Environments & Harness Design

01:23:00 - Why American Open Source Is Trailing Chinese Competitors & Trinity Adoption

01:29:25 - Hiring at Arcee & Advice to Your 20-Year-Old Self