In reality, there aren't many Open Source models on the market (which imply the openness of the entire infrastructure, including training data).
The models that many mistakenly call Open Source are actually more like free software distributions, but without the code.
The main downside of Open Weights models is that they can be fine-tuned, but fine-tuning improves response results in one area while simultaneously worsening responses in another (where the model previously performed well before fine-tuning).
In an ideal world, you want a model that responds well across the entire spectrum.