Open weight et open source, ce qui les sépare
Two labels used interchangeably even though they do not release the same things. **Open weight**: you get the model **weights**, the numerical parameters learned during training. With them you can run inference on your own machine, quantize the model to fit it in memory, fine-tune it on your data. A licence usually ships alongside the weights and sets what you are allowed to do with them. **Open source**: one notch above. Public are the weights, **the code** used for both training and inference, and **the training data**. The practical consequence: anyone could retrain the model from scratch, given the compute. One question settles it: can I **reproduce** this model, or only **run** it? Reproducing is open source. Running, quantizing, fine-tuning is open weight.
Strengths
- Verifiable in a minute: look for the training code repository and the dataset, not the wording in the announcement
- Open weight already covers local inference, quantization and fine-tuning, which is most of what real projects need
- Naming it correctly avoids promising a client an auditability the model does not allow
Limitations
- The open weight label says nothing about the licence, which can be more restrictive than a classic free licence
- A genuinely open source model stays practically unusable without the compute to retrain it
Best for
- Settling a team debate on what may be claimed in a commercial proposal
- Picking a local model knowing what the licence allows before depending on it