Open Weights, Open Source, and Who Actually Owns Your Data
Meta shipped a new open-weight model family this week, and the interesting part isn't the competitive story dominating the cable news coverage. It's that a genuinely capable model you can download and run yourself is now a real option, and that changes where your data has to travel to get value out of AI.
Mark Zuckerberg published a 14-page essay this week, and Meta shipped a new open weight model family called Muse Glimmer, sized to run on a Mac or PC with a single graphics card. He also said Meta will open the weights for Muse Spark 1.2, its most capable model, and that larger open models are coming behind it. Squawk Box spent real airtime on it that morning, mostly framed as a competitive story: Meta against the closed labs, and American labs against the Chinese open weight releases that have been setting the pace.
That framing is fine for the market. It is not the interesting part if you run an enterprise. The interesting part is that a genuinely capable model you can download and run yourself is now a real option, and that changes where your data has to travel to get value out of AI.
We have been making a version of this argument in geospatial for years. The noun changed. The argument did not.
Open weight is not the same as open source
Precision matters here, because the two terms are already getting used interchangeably and they are not the same thing.
Open Weight
The trained parameters are published. You can download them, run inference on hardware you control, fine tune on your own data, and quantize the model to fit what you actually have. What you usually do not get is the training data or the training code, and the license often carries use restrictions a closed model contract would not.
Open Source
GeoServer, PostGIS, GDAL, MapStore2, and CesiumJS give you full source under recognized licenses. You can read every line, patch it, fork it, and ship it.
Nothing has to leave your boundary for the system to work — that is the property that matters most to a security officer, and open weight and open source both deliver it.
The concern buyers are actually voicing
Alex Karp made this case on Squawk Box back in July. The delivery got more attention than the substance, and we have no interest in relitigating any of that. Strip out the theatrics and the vendor positioning and the concern underneath is one we hear directly from buyers on both the commercial and the government side.
"What you are sending is not really data, it is your process: how you price, how you triage, how you decide, how you route work."
You pay by the token. To get an answer you send proprietary context to an endpoint you do not control. Across enough calls, that is your business described in detail to a third party. Whether or not any particular provider trains on it, the exposure is structural rather than contractual.
That is not a novel AI problem. It is the same data custody and lock-in problem enterprises have been working through in every other layer of the stack, arriving in a new layer.
What we wrote last year still holds
In June of last year we published Why Open Source GIS Is Ready for Mission-Critical Government Use. The argument was that GeoServer, PostGIS, GDAL, and CesiumJS are not fringe tools waiting for permission. They are already carrying command and control, imagery processing and dissemination, and mission planning systems today. We made a version of the same argument at SOF Week, where the appetite for getting out from under vendor-locked stacks was not subtle.
The reason those customers choose open source is rarely the license fee. It is control:
Deploy
Inside a boundary you have already accredited.
Patch
On your schedule, instead of a vendor's.
Inspect
The code, any time an auditor asks.
Swap
A component the moment a better one shows up.
Open weight models hand you those same four properties on the model layer. That is why this week matters more than the horse race does.
What sovereignty looks like in practice
The mental shift is simple. Instead of moving your data to the model, you move the model to your data. Concretely, for an enterprise that gets this right:
Runs inside your boundary
Inference runs inside your own VPC, your own datacenter, or a fully disconnected enclave. There is no egress path for prompt content — nothing to negotiate in a data processing agreement, nothing to explain to a regulator.
Cost becomes capacity planning
You are buying or renting GPUs on a schedule you set. Harder to get right at the start, and far more predictable to live with.
The model is replaceable
When something better lands in eight weeks, you swap the weights and rerun your evaluations. You are not migrating a platform.
Fine-tune on data you'd never send anywhere
This is the one that moves the needle for most clients we talk to — the proprietary corpus that would make a model genuinely useful is exactly the corpus legal will not let out the door.
We are working with clients standing this up right now. We are not going to name them, but the pattern that keeps working is right-sizing rather than maximalism. Classification, extraction, summarization, routing, metadata enrichment, and geospatial attribute cleanup do not need a frontier model. They need a competent small model sitting close to the data, running constantly, cheaply, and privately. Save the frontier calls for the genuinely hard reasoning, and make that choice deliberately per workload instead of by default.
The part that does not make it onto the slide
Open source is not free, and open weight models are not free either. You are trading a license line item for engineering discipline.
On the open source geospatial side that means a real DevSecOps pipeline: software bill of materials generation, CVE tracking, a patch cadence you actually hold to, and an upgrade path you have rehearsed. NGS exists in large part because organizations want those properties without standing up a team to maintain them.
On the model side it is the same shape of work in a new place. You own it all:
Evaluation
Nobody is grading your outputs for you.
Guardrails
And prompt-injection defense, yours to build.
GPU Capacity
And utilization, yours to manage.
Upgrade Cadence
And the regression testing it demands.
Anyone telling you open weight models are simply cheaper is selling something. The accurate version is that the cost moves from a vendor invoice onto your own engineering, and in exchange you keep control of your data and your process. For a lot of enterprises that trade is obviously worth making. For some it is not. Working out which one you are is a real analysis, not a slogan, and it is most of what our advisory work consists of.
The pattern
Strip away the model names, the essay, and the cable news segment, and the same principle sits underneath all of it. Build on components you can inspect. Deploy them inside a boundary you control. Run them through a pipeline you own. Keep the option to replace any piece of it.
That is what CoreSpatial is, and it is why we built it on open source instead of wrapping somebody else's proprietary engine. It is also the right way to think about bringing AI into an enterprise. The stack changed. The principle did not.
Working through where open weight models fit in your environment?
This is work we do.