Paying Twice for Intelligence: The Open Source Moment and Our Bet on Arcee
There’s been a flurry of open source activity recently, punctuated by Moonshot AI’s release of Kimi K3, a 2.8 trillion parameter open weight model that benchmarks alongside the best closed systems from OpenAI and Anthropic. The gap between open and closed models used to be measured in years. Then it was months. Now it is weeks.
We bet on this happening a few years ago when we invested in Arcee’s seed (and Series A), and I want to explain why I think the open source moment is just getting started.
Why Open Will Win
Microsoft’s Satya Nadella recently articulated what he calls the Reverse Information Paradox. Today, the buyer of intelligence (i.e., a company or individual) risks giving away their knowledge just to use what they bought. As Nadella puts it, you pay for intelligence twice — once with money, and again with the proprietary knowledge you must reveal to make the model useful. Every prompt, every correction, every eval is institutional know-how leaking out, trace by trace, to a vendor who may someday compete with you. His conclusion: enterprises need a hard trust boundary inside which their data, evals, adapted weights, and organizational memory can compound — and the right to fine-tune and train their own models within it.
Open source is a smart and differentiated business strategy to do precisely that. At Flybridge, we have been big believers and backers of ambitious open source projects for decades, including MongoDB, Appwrite, and Netbox. Among other strongly held beliefs, we believe in the power of community and decentralization.
Nadella points to another driving force behind open source in the Age of AI: the competitive paranoia among large companies about leaking alpha. Open weights let enterprises run models on their own infrastructure without leaking anything. And as Sriram Krishnan points out, it is now clear you can reach near-frontier performance with an open training lineage. With more and more powerful independent harnesses (like our portfolio company Tasklet) and evals, a company can bring their individual business context to the model of their choice.
The Arcee story
These themes are the essence of the Arcee story. When we backed the visionary Arcee team in 2023, they were a small team building custom models for enterprises by training open weight models on private data. Their models were small (so was their team), and the broader market mostly ignored them. Our view was that companies would eventually demand models they could download, inspect, and run inside their own clouds. We didn't know the exact path — in fast-moving markets, you rarely do — but we believed the direction was right and the team could navigate.
The path turned out to run through open source itself. Charles Goddard, now Arcee's Chief of Frontier Research, wrote MergeKit, the standard open source tool for merging models, and the company built its early business on top of it. Merging taught the team how to train, and training taught them how to scale.
Then this past January, they did something audacious. A small, talented team of 26 ran a 33-day training run on 2,048 NVIDIA GPUs and produced Trinity Large: a 400 billion parameter mixture-of-experts model, trained from scratch in America, released under a fully permissive Apache 2.0 license.
The major labs spend billions to build frontier models; Arcee did it for a fraction of that. CTO Lucas Atkins calls the approach "engineering through constraint," and it has earned the team enormous respect in the technical community. An amazing example of a 10x team using AI to punch far above their headcount.
This milestone was just a first step, achieved with limited capital and compute. The goal for Arcee's next generation of models is to build some of the most capable open weight systems in the world. Trinity proved what they can do under tight constraints. For now, Trinity Large Thinking is already among the most capable open weight reasoning models from any Western lab and a top choice for developers building AI agents on OpenRouter.
That "Western lab" qualifier matters. Most of the leading open models today — Kimi, DeepSeek, Qwen — come from Chinese teams. But American banks, defense contractors, and government agencies increasingly need domestic alternatives they can download, run, and audit. Trinity provides exactly that, and the Arcee team saw the need coming years before it was fashionable.
Where this goes
The frontier model providers make the case for open models every time they change prices, terms, or access overnight. As model quality converges, enterprise decisions will come down to cost, control, and trust — and on all three, open weights that live inside your own trust boundary are hard to beat. Every firm has a right to its own learning loop. The knowledge your company creates by using AI should belong to your company. Arcee built for that world before the market knew it wanted it, and Mark, Lucas, and the team are just getting started. If you want to learn more, send me a note.
