Home

Writing

Open Models, Closed Models and the Layer Above Both

3 min read

A practical distinction between open source AI, open-weight models and hosted systems - and why the user intelligence layer should remain portable.

The argument about open and closed AI is often staged like a moral choice.

One side represents freedom, transparency and local control. The other represents capability, convenience and safety. Reality is less clean.

There are at least three different categories people collapse into the word “open”:

They are not the same.

Open weights are not automatically open source.#

An open-weight model makes its trained parameters available. That can allow local inference, private deployment, fine-tuning and independence from one hosted API.

Open source AI asks for more. The Open Source Initiative definition centres the freedoms to use, study, modify and share the system, together with access to the preferred form required to make meaningful modifications. That can include code, parameters and detailed information about training data and process.

The distinction matters because control exists at several layers.

You may control the weights but not understand the data that shaped them. You may have source code but lack the compute to reproduce the system. You may use a transparent model inside an opaque product.

Openness is a stack, not a badge.

Closed models are not simply the enemy.#

Hosted models can provide frontier capability without forcing every developer to become an inference company. They absorb deployment, optimization, hardware, upgrades and much of the operational burden.

For many products, that is rational.

A small team may gain more freedom by using a strong hosted model and spending its time on the problem customers actually have. Owning weights is not useful if maintaining them consumes the company.

Closed systems become dangerous when the product’s identity, memory and data are inseparable from the provider. A model change, policy change, price change or discontinued feature can then erase the product’s foundation.

The durable architecture is model-flexible.#

Catalyst is designed around a simple assumption: no model should permanently own the relationship with the user.

Use the strongest hosted model when it is the right tool. Use an open-weight model when privacy, customization, cost or local control matters. Use several models when different tasks require different tradeoffs.

The intelligence layer above them should remain portable:

This is where continuity lives.

Open and closed systems will coexist.#

The performance gap will move back and forth. Recent AI Index data shows top providers clustering closely, while the best closed model still held a measurable lead over the best open model in early 2026. That does not produce one permanent winner. It creates a market where capability, cost, reliability, privacy and control keep trading places.

The product decision should not be ideological.

It should ask:

The model is one layer.#

Open models expand who can build. Closed models often expand what can be built immediately. Both matter.

The mistake is confusing the model with the complete system.

A useful AI product is also its context, memory, tools, evaluations, interfaces, feedback loops, security model and the relationship it develops with the person using it.

That layer should belong to the builder and, ultimately, to the user.