TPEE Toptronic® Prompt Engineering Editor

106. Model Weights 5 - Expert

Govern weights as an asset: licences, provenance, versioning, self-hosting economics and the honest limits of a checkpoint.

By Jacques Botte, founder of Toptronic®. Last updated 17 October 2026.

The lesson

Weights are a licensed asset, and the licence is not the same as the repository's. A model card may say open weights while the licence restricts commercial use, requires a name prefix, caps monthly active users, or forbids distillation into another model. Those terms bind the deployer, not the downloader, and they travel with derivatives. Read the licence file itself, record its version, and check whether an update changed it.

Provenance is a security property. A checkpoint is data that a loader will map or deserialize, and a checkpoint from an unknown source is an executable risk in every format except the strictest data-only ones. Verify published hashes, prefer signed or mirrored releases you can attribute, and pin the exact revision you deployed. Reproducibility across a model update is not automatic: the same name at a later commit is a different model.

Self-hosting economics are often miscalculated. The comparison is not the API price against zero; it is the API price against hardware amortised over its useful life, electricity, the engineer time to keep a serving stack patched, the idle capacity you must hold for peak, and the opportunity cost of that attention. Self-hosting wins on data residency, on very high sustained volume, and on tasks where a small specialised model beats a large general one. It rarely wins on a spiky low-volume workload.

Evaluation is the part that decays. A model chosen six months ago was compared against the alternatives of six months ago, using an evaluation set that has since been memorised by newer training runs. Re-run your own set against the current candidate before assuming the incumbent is still the best choice, and re-run it after every model or runtime update. The number that mattered last quarter is not evidence about this one.

Finally, be honest about what a checkpoint cannot tell you. Its weights do not document their training data, so questions of copyright, consent, bias and contamination cannot be answered by inspection. A benchmark score does not predict your workload. And a model that performs well does not thereby become accountable: the responsibility for what it produces remains with the person who deployed it, which is the same conclusion every other lesson in this course arrives at from a different direction.

Check yourself

Question 1: Which statement about an open-weights model licence is correct?
  1. It always permits unrestricted commercial use
  2. It may restrict commercial use, require attribution or cap users, and those terms travel with derivatives — correct
  3. It only applies to the downloader
  4. It is the same as the repository's code licence

Answer: It may restrict commercial use, require attribution or cap users, and those terms travel with derivatives

Weights licences commonly carry commercial, attribution and scale conditions that bind the deployer and derivatives.

Question 2: What is a genuine security concern when obtaining a checkpoint?
  1. The file is large
  2. A loader may deserialize code, so an untrusted checkpoint is an executable risk — correct
  3. The tokenizer is included
  4. The config is JSON

Answer: A loader may deserialize code, so an untrusted checkpoint is an executable risk

Data-only formats reduce but do not remove supply-chain risk; provenance and hash verification are the controls.

Question 3: Which is a fair statement about what weights can tell you?
  1. They document their training data
  2. They prove the model is unbiased
  3. They do not reveal their training data, so copyright, consent and contamination questions cannot be answered from them — correct
  4. They guarantee benchmark performance on your task

Answer: They do not reveal their training data, so copyright, consent and contamination questions cannot be answered from them

A checkpoint is opaque about its data. Claims about provenance and fairness need evidence outside the file.

← Previous lesson · All 106 lessons

The full course — 106 lessons and 318 quiz questions — ships inside the app. Get TPEE to study it offline.