What weights are for, after inference moves on
Most conversations about open weights are about inference: run it locally, serve it cheaper, avoid rate limits. Useful — but it frames preservation as nostalgia, keeping old things because you liked them. The stronger argument points the other way: weights are construction material, and the constructions that matter most happen after the artifact stops being the current fashion.
Hold a full-precision copy of a base model and you hold the starting point for work nobody has thought of yet. Fine-tuning takes a general artifact and specializes it — your data, your domain, your voice. Distillation compresses a capability into a smaller student that runs where the original cannot. Merging and layer surgery compose pieces of different models into genuinely new objects. Every one of these operations requires the tensors themselves, at useful fidelity, on storage you control. None of them can be performed against an API.
That last clause is the crux. A hosted endpoint lets you use a capability; provider-side fine-tuning exists, but what it produces stays on the provider's infrastructure — you rent the result, you never hold the modified artifact. The moment your project is "take this model and make it mine" — mine in tone, mine in domain, mine in footprint — you need the bytes. And if the provider has since gated, degraded, or retired the distribution, the project's starting material is gone regardless of how willing you are to pay.
The vessel framing
A vessel here is any future system you might build whose irreplaceable input is a model you can no longer assume will be fetchable: an on-device assistant distilled from a teacher that has since been pulled; a domain adapter whose base was re-released in a lossy format; a research replication that needs the exact artifact a paper used. You cannot enumerate tomorrow's vessels — that is the point of raw material — but you can keep the pile that makes them buildable.
Fidelity is what makes the argument bite. You cannot fine-tune your way back to information a lossy copy already discarded; quantization is a one-way door for the artifacts that pass through it. That is why an acquisition policy that optimizes for the highest-fidelity authoritative representation — rather than the smallest runnable file or the biggest blob on the page — is the correct default for a woodpile, and why quantized derivatives earn their place only as deliberate additions, clearly labeled as what they are.
Keeping the option open
You do not need a concrete vessel plan to justify the pile — you need the recognition that plans of that shape exist and are common in hindsight. The cheap strategy is asymmetric: acquiring one authoritative copy of each model you respect costs disk and patience; discovering in two years that the necessary base is unfetchable costs the project. Optimize for keeping doors open, and let future-you choose which to walk through.