Jesse Middleton@srcasm
The gap between open and closed AI models used to be measured in years. Then it was months. Now it is weeks.
Moonshot AI recently released Kimi K3, a 2.8 trillion parameter open weight model that benchmarks alongside the top closed systems. We bet on this shift a few years ago when we invested in @arcee_ai, and the open source moment is just getting started.
@satyanadella recently highlighted what he calls the Reverse Information Paradox. Today, buyers of AI risk giving away their knowledge just to use the software.
You pay for intelligence twice. You pay once with money, and again with the proprietary knowledge you must reveal to make the model useful.
Every prompt and evaluation is institutional know-how leaking to an outside vendor. Enterprises need a hard trust boundary where data, evaluations, and organizational memory can compound safely. They need the right to fine-tune and train their own models within that boundary.
Independent tools from companies like @TaskletAI let businesses bring their unique context to any model they choose. This makes open source a smart enterprise strategy. At @flybridge, we have backed open source projects like @MongoDB, @appwrite, and @netboxlabs for decades. Open weights let companies run models on their own infrastructure without leaking their competitive edge.
When we backed Arcee in 2023, they were a small team building custom models by training open weight systems on private data. We believed companies would eventually demand models they could download, inspect, and run inside their own clouds.
This past January, their team of 26 ran a 33-day training run on 2,048 GPUs and produced Trinity Large. It is a 400 billion parameter mixture-of-experts model, trained from scratch in America and released under an Apache 2.0 license.
The major labs spend billions to build frontier models. Arcee did it for a fraction of that through constrained engineering. Trinity Large Thinking is now a top choice for developers building AI agents on @OpenRouter.
Most leading open models today come from Chinese teams. American banks, defense contractors, and government agencies increasingly need domestic alternatives they can download, run, and audit. Trinity provides exactly that.
As model quality converges, enterprise decisions will come down to cost, control, and trust. Open weights that live inside your own trust boundary win on all three.
If you are currently evaluating/building open source AI tools, send me a note.
I’d love to connect.
