Open Weights and Proprietary AI Muscle Memory

Open-weight models are getting more and more attention lately. I’ve posted about this trend before, but I think it’s worth digging a little deeper into what it actually means.

First, a quick clarification.

For the most part, these are not actually open source models (for example, Kimi K3). That’s a small but important distinction. Open source would mean releasing the full recipe to recreate the model: the training code, methodology, datasets (or instructions to recreate them), architecture, and weights. Most of today’s “open” models don’t do that.

A more accurate description is open-weight models.

Open-weight models provide the trained weights so you can run them yourself, fine-tune them, quantize them, and deploy them wherever you choose.

The interesting part is the fine-tuning.

Think of the weights as AI’s muscle memory. Prompts tell the model what you want to do today. The weights determine what comes naturally.

Frontier models are trained on an enormous breadth of knowledge, making them incredibly capable generalists. But they’re designed to work reasonably well for almost everyone, not exceptionally well for your specific business.

That’s why we wrap them with retrieval systems (RAG), system prompts, tools, and other techniques to inject business-specific knowledge and behavior.

With an open-weight model, some of that expertise can move into the model itself. Instead of repeatedly explaining your terminology, workflows, or preferred outputs through prompts and context, you can fine-tune the model so those patterns become part of its muscle memory. For many repeated workflows, that means less prompt engineering, less context injection, and often lower inference costs.

Open-weight models aren’t “free”. You still have to host and operate them, but they unlock something far more interesting than lower API costs.

They make proprietary AI muscle memory possible.

I think the next wave may be proprietary weights. Models that encode a company’s unique expertise, language, workflows, and decision-making.

The companies that own that muscle memory may ultimately have an advantage that goes well beyond simply choosing the best frontier model.