THE VERDICT
● Self-hosted open weights if you need to audit what runs, or data cannot leave your infrastructure.
● Hosted API if you need someone contractually accountable when it goes wrong.
● What changed: nothing technical. What you have to be able to prove.
The two positions
| Self-hosted weights | Hosted API |
|---|
| Can you audit what runs | Yes | No |
| Data leaves your infrastructure | No | Yes |
| Someone accountable if it fails | You | A contract |
| Safety behaviour guaranteed | By your own testing | By the vendor, contractually |
| Can disappear on you | No. You hold the weights | Yes. Deprecations happen |
| Cost shape | Hardware, fixed | Per token, variable |
THE ROW THAT MOVED IS THE FOURTH
Self-hosting always meant safety behaviour was your responsibility to verify. Most teams treated it as inherited from the lab that published the weights.
With modified versions now sold commercially, "we use the official weights" becomes a claim you have to be able to evidence rather than an assumption. Checksum what you downloaded, record where it came from, and be able to show it.
What to do either way
- Record the provenance of every model you run. Source, date, version, checksum. If you cannot produce that, you cannot answer the question when it comes.
- Test safety behaviour yourself if you self-host. It was always your job; now someone may ask you to prove you did it.
- Read what a hosted vendor actually warrants. Accountability is only worth what the contract says.
- Keep a fallback either way. Deprecations, licence changes and advisories have all happened this quarter.
Which one
| If you are... | Pick |
|---|
| Handling data that cannot leave | Self-hosted, with documented provenance |
| Selling into regulated sectors | Hosted, for the accountability. Or both, documented |
| Cost-constrained at high volume | Self-hosted. Fixed cost wins above a threshold |
| Small team, variable usage | Hosted. Hardware you use half the time is waste |
FAQ
Does the stripped-model story affect my self-hosted setup?
Not operationally. Weights you downloaded are unchanged. What changed is that "we use the official weights" is now a claim worth being able to evidence.
Is self-hosting less safe now?
No. The same weights behave the same way. The category has a documented downstream, which is a procurement question rather than a technical one.
What should I record?
Source, date, version and checksum for every model you run. That is the whole answer to the question when someone asks it.