The semantics matter because they determine who builds the next economy.
The Summary
- LLMs fall into three distinct categories: proprietary (GPT, Claude, Gemini), open weight (Meta's Llama), and open source (fully auditable code and training data).
- Open weight models publish their parameters but not training code or data, creating a middle ground that's widely misunderstood as "open source."
- The distinction determines who can build agents, who can audit safety, and who controls the infrastructure of Web4.
The Signal
Proprietary models run through cloud APIs. You send your prompt to OpenAI's servers, they run inference, you get a response. The model weights, training data, and architecture stay locked in their data centers. This makes them dead simple to integrate. Businesses plug in via API and ship features in days, not months. The trade-off is total dependency. If OpenAI changes pricing, rate limits, or terms of service, you adapt or rebuild.
The frontier advantage is real. GPT and Claude lead on reasoning tasks because they trained on orders of magnitude more compute and data than anyone else can afford. That gap isn't closing for most companies. But "frontier" is a moving target, and the delta between proprietary and open weight models shrinks monthly.
"Open weight models publish their parameters but not training code or data, creating a middle ground that's widely misunderstood."
Open weight sits in the middle. Meta's Llama series is the flagship example. You can download the model weights and run inference locally or on your own infrastructure. No API dependency. No per-token fees after initial compute costs. But you can't see how it was trained. The training code, data mix, and fine-tuning methods stay private. You get the finished product, not the recipe.
This matters for three reasons:
- Cost structure shifts from operational to capital. Running Llama locally means upfront GPU spend, but zero marginal cost per query after that.
- Data sovereignty. Your prompts never leave your infrastructure. Critical for regulated industries or anyone building agents that touch sensitive data.
- Customization limits. You can fine-tune on your own data, but you can't audit what biases or vulnerabilities came baked into the base model.
The "open weight" label does heavy marketing work. It sounds like open source. It gets lumped in with open source in policy debates. But it's not the same thing. True open source models publish everything: weights, training code, data sources, evaluation benchmarks. You can reproduce the model from scratch, audit every decision, and fork it if you disagree with the direction.
The Implication
For builders, the choice is strategic. Proprietary makes sense when you need bleeding-edge capabilities and can stomach vendor lock-in. Open weight works when you need cost control, data privacy, or the ability to run agents that don't phone home. True open source is rare but critical for anyone building infrastructure that needs to be auditable or forkable.
The next 24 months will clarify which category wins in the agent economy. If agents become commodity infrastructure, open weight wins. If reasoning capabilities keep widening the frontier gap, proprietary stays dominant. Watch which model type the top 100 AI companies standardize on. That tells you who controls Web4.